Systems and methods for mental health assessment

Abstract

The present disclosure provides systems and methods for assessing a subject's mental state in a single session or across multiple different sessions, using, for example, an automated module, and presenting and / or constructing at least one query based in part on one or more target mental states to be assessed. The query may be configured to elicit at least one response from the subject. The query may be transmitted to the subject in audio, visual, and / or text format to elicit a response. Data including the response from the subject may be received. The data may be processed using one or more individual, combined, or fusion models. One or more assessments of the mental state associated with the subject may be generated for a single session, for each of multiple different sessions, or upon completion of one or more of the multiple different sessions. [Selection diagram] Figure 20B

Description

[Technical field]

[0001] cross reference This application is a continuation of U.S. Provisional Application No. 62 / 687,176, filed June 19, 2018, U.S. Provisional Application No. 62 / 749,113, filed October 22, 2018, U.S. Provisional Application No. 62 / 749,654, filed October 23, 2018, U.S. Provisional Application No. 62 / 749,663, filed October 23, 2018, U.S. Provisional Application No. 62 / 749,669, filed October 23, 2018, U.S. Provisional Application No. 62 / 749,672, filed October 24, 2018, U.S. Provisional Application No. 62 / 754,534, filed November 1, 2018, and U.S. Provisional Application No. 62 / 749,673, filed October 24, 2018. This application claims priority to U.S. Provisional Application No. 62 / 754,541, filed November 1, 2018, U.S. Provisional Application No. 62 / 754,547, filed November 1, 2018, U.S. Provisional Application No. 62 / 755,356, filed November 2, 2018, U.S. Provisional Application No. 62 / 755,361, filed November 2, 2018, U.S. Provisional Application No. 62 / 733,568, filed September 19, 2018, and U.S. Provisional Application No. 62 / 733,552, filed September 19, 2018, each of which is fully incorporated by reference in its entirety for all purposes. [Background technology]

[0002] Behavioral health is a serious issue. In the United States, suicide ranks among the top 10 causes of death as reported by the Centers for Disease Control (CDC). According to the World Health Organization (WHO), depression is the leading cause of disability worldwide. Screening for depression and other mental health disorders by physicians and health care providers is widely recommended. The current "gold standard" for screening or monitoring patients for depression is the Patient Health Questionnaire 9 (PHQ-9), a written depression screening or monitoring test with nine (9) multiple-choice questions. Other similar assessment tests include the PHQ-2 and the Generalized Anxiety Disorder 7 (GAD-7).

[0003] Many believe that the PHQ-9 and other similar screening or monitoring tools for detecting behavioral health conditions such as depression are inadequate. The PHQ-9 is said to successfully detect depression in 85-95% of patients, but it is also said that 54% of all suicides are committed by people without a diagnosis of depression. These two claims seem to be completely contradictory to each other in terms of screening or monitoring, but in reality, not enough people are screened. Summary of the Invention [Problem to be solved by the invention]

[0004] Part of the problem is that traditional screening or monitoring surveys are uninteresting due to their repetitive nature and lack of personalization. Another problem is that patients may be dishonest in their responses to assessment tools, and the PHQ-9 and similar tools do not provide a mechanism by which dishonesty in patients' responses can be assessed. Finally, these surveys require effort on the part of clinicians and patients, as some patients require assistance in completing the survey, which disrupts both clinician and patient workflow. [Means for solving the problem]

[0005] The present disclosure provides systems and methods that can more accurately and effectively assess, screen, estimate, and / or monitor the mental state of a human subject when compared to traditional mental health assessment tools. In one aspect, a method is provided for assessing a mental state of a subject in a single session or across multiple different sessions. The method can include using an automated module to present and / or construct at least one query based in part on one or more target mental states to be assessed. The at least one query can be configured to elicit at least one response from the subject. The method can also include sending the at least one query in audio, visual, and / or text format to the subject to elicit at least one response. The method can also include receiving data including at least one response from the subject in response to sending the at least one query. The data can include speech data. The method can further include processing the data using one or more individual, combined, or fusion models including a natural language processing (NLP) model, an acoustic model, and / or a visual model. The method may further include generating one or more assessments of a mental state associated with the subject for a single session, for each of a plurality of different sessions, or upon completion of one or more of the plurality of different sessions.

[0006] In some embodiments, one or more of the individual models, combined models, or fusion models may include a metadata model that may be configured to use demographic information and / or medical history of a subject to generate one or more assessments of a mental condition associated with the subject.

[0007] In some embodiments, at least one query can include multiple queries and at least one response can include multiple responses. The multiple queries can be sent to the subject sequentially, and the multiple queries can be configured to systematically elicit multiple responses from the subject. In some embodiments, the multiple queries can be structured hierarchically, such that each subsequent query in the multiple queries is structured as a logical follow-on to the subject's response to a preceding query and can be designed to assess or draw inferences regarding multiple aspects of the subject's mental state.

[0008] In some embodiments, the automated module may be further configured to submit and / or construct at least one query based in part on the subject's profile.

[0009] In some embodiments, the one or more target mental conditions may be selected from the group consisting of depression, anxiety, post-traumatic stress disorder (PTSD), schizophrenia, suicidality, and bipolar disorder.

[0010] In some embodiments, the one or more target mental conditions can include one or more conditions or disorders associated with or coexisting with a predefined list of mental disorders. The predefined list of mental disorders can include mental disorders defined or provided in the Diagnostic and Statistical Manual of Mental Disorders. In some embodiments, the one or more associated or coexisting conditions or disorders can include fatigue, loneliness, lack of motivation, or stress.

[0011] In some embodiments, the assessment can include a score indicating whether the subject (i) is more likely than others to experience at least one of the target mental states or (ii) is more likely than others to experience at least one of the target mental states at some time in the future, in some embodiments, the future time can be within a clinically actionable future range.

[0012] In some embodiments, the method can further include transmitting the assessment to medical personnel for use in assessing the mental state of the subject, which can be performed during the assessment, just in time, or in real time after the assessment is completed.

[0013] In some embodiments, multiple queries may be designed to test or detect multiple aspects of a subject's mental state.

[0014] In some embodiments, the assessment may include a score indicating whether the subject (i) is more likely than others to experience at least one of the target mental states or (ii) is more likely than others to experience at least one of the target mental states at some point in the future. The score may be calculated based on processed data obtained from a plurality of responses of the subject to a plurality of queries. In some embodiments, the score may be continuously updated with processed data obtained from each of the subject's subsequent responses to a preceding query.

[0015] In some embodiments, the method may further include identifying further information to be elicited from the subject based on at least one response. The method may further include sending a subsequent query to the subject. The subsequent query may be related to the further information and configured to elicit a subsequent response from the subject. The method may also include receiving data including the subsequent response from the subject in response to sending the subsequent query. The method may further include processing the subsequent response to update the assessment of the mental state of the subject. In some embodiments, identifying further information to be elicited from the subject may include identifying (i) one or more elements of substantial content or (ii) one or more patterns in the data that are significant to the mental state of the subject. The method may further include identifying one or more items of follow-up information related to the one or more elements or one or more patterns to be asked of the subject for each of the one or more elements or one or more patterns of substantial content, and generating a subsequent query. The subsequent query may be related to the one or more items of follow-up information.

[0016] In some embodiments, the NLP model may be selected from the group consisting of an emotion model, a statistical language model, a topic model, a syntax model, an embedding model, a dialogue or discourse model, an emotion or affect model, and a speaker personality model.

[0017] In some embodiments, the data may further include an image or video of the subject. The data may be further processed using a vision model to generate an assessment of the subject's mental state. In some embodiments, the vision model may be selected from the group consisting of a facial cue model, a body movement / motion model, and an eye activity model.

[0018] In some embodiments, the at least one query may be sent in a conversational context in the form of a question, statement, or comment configured to elicit at least one response from the subject. In some embodiments, the conversational context may be designed to encourage elicitation of a genuine, reflective, thoughtful, or candid response from the subject. In some embodiments, the conversational context may be designed to influence the time it takes the subject to compose at least one response. In some embodiments, the method may further include sending one or more prompts in audio and / or visual form to the subject when a latency threshold is exceeded. In some embodiments, the conversational context may be designed to enhance one or more performance metrics of the assessment of the subject's mental state. In some embodiments, the one or more performance metrics may be selected from the group consisting of F1 score, area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), and equal error rate.

[0019] In some embodiments, the at least one query is not or need not be sent or given in the form of a standardized test or questionnaire. In some embodiments, the at least one query may include subject matter adapted or modified from a standardized test or questionnaire. In some embodiments, the standardized test or questionnaire may be selected from the group consisting of PHQ-9, GAD-7, HAM-D, and BDI. The standardized test or questionnaire may be other similar tests or questionnaires for assessing the mental health status of a patient.

[0020] In some embodiments, one or more of the individual models, the combined model, or the fusion model may include a regression model.

[0021] In some embodiments, the at least one query may be designed to be non-restrictive, without restricting the at least one response from the subject to being a binary yes or no response.

[0022] In some embodiments, the scores can be used to calculate one or more scores with clinical values.

[0023] In some embodiments, the assessment may include a quantized score estimate of the subject's mental state. In some embodiments, the quantized score estimate may include a calibrated score estimate. In some embodiments, the quantized score estimate may include a binary score estimate.

[0024] In some embodiments, multiple queries can be represented as a series of edges, and multiple responses can be represented as a series of nodes in a node network.

[0025] In some embodiments, the mental condition may include one or more medical, psychological, or psychiatric conditions or symptoms.

[0026] In some embodiments, the method may be configured to further evaluate a physical state of the subject revealed based on the speech data of the subject. The method may further include processing the data using one or more individual, combined, or fusion models to generate an evaluation of the physical state of the subject. In some embodiments, the physical state may include a score indicating whether the subject (i) is more likely than others to experience at least one of the plurality of physiological states or (ii) is more likely than others to experience at least one of the physiological states at some point in the future.

[0027] In some embodiments, the subject's physical condition is manifested due to one or more physical conditions that affect the characteristics or quality of the subject's voice.

[0028] In some embodiments, the automated module may be a mental health screening module that may be configured to dynamically construct at least one query based in part on one or more target mental states to be assessed.

[0029] In some embodiments, one or more of the individual models, combined models, or fusion models may include a composite model, which may be an aggregation of two or more different models.

[0030] Another aspect of the present disclosure provides a non-transitory computer-readable medium comprising machine-executable instructions that, when executed by one or more computer processors, perform any of the foregoing methods described above or elsewhere herein.

[0031] Another aspect of the present disclosure provides a system comprising one or more computer processors and a memory containing machine-executable instructions that, when executed by the one or more computer processors, perform any of the methods described above or elsewhere herein.

[0032] Another aspect of the present disclosure provides a method for screening or monitoring a subject or diagnosing a subject for a mental health disorder. The method may include transmitting at least one query to the subject. The at least one query may be configured to elicit at least one response from the subject. The method may further include receiving data including the at least one response from the subject in response to transmitting the at least one query. The data may include speech data. The method may further include processing the data using one or more individual, combined, or fusion models including a natural language processing (NLP) model, an acoustic model, and / or a visual model to generate an output. The method may further include generating a score and a confidence level for the score using at least the output. The score may include a prediction that the subject has a mental health disorder. The confidence level may be based at least in part on a quality of the speech data and represents the degree to which the prediction can be trusted.

[0033] In some embodiments, one or more of the individual models, combined models, or fusion models can include a metadata model that can be configured to use demographic information and / or medical history of a subject to generate one or more assessments of a mental condition associated with the subject.

[0034] In some embodiments, the output may include an NLP output, an acoustic output, and a visual output. In some embodiments, the NLP output, the acoustic output, and the visual output may each include multiple outputs corresponding to different time ranges of the data. In some embodiments, generating the score may include (i) dividing the NLP output, the acoustic output, and the visual output into discrete time segments, (ii) assigning a weight to each discrete time segment, and (iii) calculating a weighted average of the NLP output, the acoustic output, and the visual output using the assigned weights. In some embodiments, the weights may be based on at least (i) base weights of one or more individual, combined, or fused models, and (ii) a confidence level of each discrete time segment of the NLP output, the acoustic output, and the visual output.

[0035] In some embodiments, one or more individual models, combined models, or fusion models may be interdependent such that each of the one or more individual models, combined models, or fusion models is conditional on the output of at least one other of the one or more individual models, combined models, or fusion models.

[0036] In some embodiments, generating the score may include fusing the NLP output, the audio output, and the visual output.

[0037] In some embodiments, generating a confidence level for the score may include fusing (i) a confidence level of the NLP output with (ii) a confidence level of the acoustic output.

[0038] In some embodiments, the method can further include converting the score into one or more scores with clinical values.

[0039] In some embodiments, the method may further include transmitting the one or more scores with clinical values ​​to the subject and / or the subject's contacts. In some embodiments, the method may further include transmitting the one or more scores with clinical values ​​to a healthcare professional for use in evaluating the subject's mental health and / or in providing mental health care. In some embodiments, transmitting may include transmitting the one or more scores with clinical values ​​to a healthcare professional during screening, monitoring, or diagnosis. In some embodiments, transmitting may include transmitting the one or more scores with clinical values ​​to a healthcare professional or payer after the screening, monitoring, or diagnosis is completed.

[0040] In some embodiments, the at least one query can include multiple queries and the at least one response can include multiple responses. Generating the score can include updating the score after receiving each of the multiple responses, and the method can further include converting the score to one or more scores with clinical values ​​after each update. The method can further include transmitting the one or more scores with clinical values ​​to a health care professional after conversion.

[0041] In some embodiments, the method may further include determining that the confidence level does not meet the predetermined criterion, generating, in real time, at least one further query based at least in part on the at least one response, and repeating steps (a)-(d) using the at least one further query until the confidence level meets the predetermined criterion.

[0042] In some embodiments, the confidence level can be based on the length of at least one response. In some embodiments, the confidence level can be based on the assessed veracity of one or more responses of the subject.

[0043] In some embodiments, one or more individual, combined, or fusion models can be trained on speech data from a plurality of subjects, each of which has completed a survey or questionnaire indicating whether the subject has a mental health disorder, and the confidence level can be based on the veracity of responses assessed in the survey or questionnaire.

[0044] In some embodiments, the method may further include extracting one or more topics of interest of the subject from the speech data using a topic model.

[0045] In some embodiments, the method can further include generating a word cloud from the one or more topics of interest. The word cloud reflects changes in the one or more topics of interest of the subject over time. In some embodiments, the method can further include transmitting the one or more topics of interest to a health care professional, the subject, or both.

[0046] In some embodiments, the video output may be assigned a higher weight than the NLP output and audio output in generating the score when the subject is not speaking, in some embodiments, the weight of the video output in generating the score may be increased when the NLP output and audio output indicate that the subject's veracity level is below a threshold.

[0047] In some embodiments, the video model may include one or more of a facial cue model, a body movement / motion model, and a gaze model.

[0048] In some embodiments, at least one query can include multiple queries and at least one response can include multiple responses. The multiple queries can be configured to sequentially and systematically elicit multiple responses from the subject. The multiple queries can be hierarchically structured such that each subsequent query in the multiple queries can be a logical follow-on to the subject's response to a preceding query and can be designed to assess or draw inferences regarding different aspects of the subject's mental state.

[0049] In some embodiments, at least one query may include a subject matter adapted or modified from a clinically validated survey, test, or questionnaire.

[0050] In some embodiments, the acoustic model may include one or more of an acoustic embedding model, a spectro-temporal model, a hypervector model, an acoustic influence model, a speaker personality model, an intonation model, a speaking rate model, a pronunciation model, a non-verbal model, or a fluency model.

[0051] In some embodiments, the NLP models may include one or more of an emotion model, a statistical language model, a topic model, a syntax model, an embedding model, a dialogue or discourse model, an emotion or sentiment model, or a speaker personality model.

[0052] In some embodiments, the mental health disorder can include depression, anxiety, post-traumatic stress disorder, bipolar disorder, suicidal tendencies, or schizophrenia.

[0053] In some embodiments, the mental health disorder can include one or more medical, psychological, or psychiatric conditions or symptoms.

[0054] In some embodiments, the score may include a score selected from a range, which may be normalized for a general population or a particular population of subjects.

[0055] In some embodiments, the one or more scores with a clinical value may include one or more descriptors associated with a mental health disorder.

[0056] In some embodiments, the above steps (a)-(d) may be repeated at multiple different times to generate multiple scores. The method may further include transmitting the multiple scores and confidences to a computing device, and displaying, on the computing device, the multiple scores and confidences graphically as a function of time on a dashboard or other display for one or more end users.

[0057] In some embodiments, the quality of the speech data may include a quality of an audio signal of the speech data.

[0058] In some embodiments, the quality of the speech data may include an indication of the confidence of a speech recognition process performed on the audio signals of the speech data.

[0059] In some embodiments, the method may be performed with respect to a single session, and the score and confidence level of the score may be generated with respect to the single session.

[0060] In some embodiments, the method may be performed for and across a number of different sessions, and a score and a confidence level for the score may be generated for each of the multiple different sessions or upon completion of one or more of the multiple different sessions.

[0061] Another aspect of the present disclosure provides a non-transitory computer-readable medium that includes machine-executable instructions that, when executed by one or more computer processors, perform any of the methods described above or elsewhere herein.

[0062] Another aspect of the present disclosure provides a system comprising one or more computer processors and a memory containing machine-executable instructions that, when executed by the one or more computer processors, perform any of the methods described above or elsewhere herein.

[0063] Another aspect of the present disclosure provides a method for processing speech and / or video data of a subject to identify a mental state of the subject. The method may include receiving speech and / or video data of the subject, processing the speech and / or video data using at least one processing technique to identify the mental state with (i) a low error rate that is at least 10% lower, or (ii) with at least 10% greater accuracy than a standardized mental health questionnaire or testing tool that can be used to identify the mental state. The low error rate or accuracy may be established relative to at least one or more benchmark standards that can be used by an entity to identify or assess one or more medical conditions that constitute the mental state.

[0064] In some embodiments, the entity may include one or more of the following: a clinician, a medical professional, an insurance company, and a government regulatory agency. In some embodiments, the at least one or more benchmark criteria may include at least one clinical diagnosis that has been independently verified to be accurate in identifying a mental condition. In some embodiments, the speech data may be received in near real-time as the subject is speaking. In some embodiments, the speech data may be generated in an offline mode from a stored recording of the subject's speech.

[0065] Another aspect of the present disclosure provides a method for processing speech data of a subject to identify a mental state of the subject. The method may include receiving speech data of the subject, and processing the speech data using at least one processing technique to identify a mental state, wherein the identification of the mental state is according to one or more performance metrics compared favorably to a standardized mental health questionnaire or testing tool usable to identify the mental state.

[0066] In some embodiments, the one or more performance metrics may include sensitivity or specificity, and the speech data may be processed according to a desired level of sensitivity or a desired level of specificity. In some embodiments, the desired level of sensitivity or a desired level of specificity may be defined based on criteria established by an entity. In some embodiments, the entity may include one or more of the following: clinicians, medical professionals, insurance companies, and government regulatory agencies.

[0067] Another aspect of the present disclosure provides a method for processing speech data of a subject to identify or assess a mental state of the subject. The method may include receiving speech data of the subject, processing the speech data using one or more processing techniques to generate one or more descriptors indicative of the mental state, and generating a plurality of visual elements of the one or more descriptors. The plurality of visual elements may be configured to be displayed in a graphical user interface of the user's electronic device for use by a user to identify or assess the mental state.

[0068] In some embodiments, the user may be a subject. In some embodiments, the user may be a clinician or medical professional. In some embodiments, the one or more descriptors may include a calibrated or normalized score indicative of the mental state. In some embodiments, the one or more descriptors may further include a confidence level associated with the calibrated or normalized score.

[0069] Another aspect of the present disclosure provides a method for identifying, assessing, or monitoring a mental state of a subject. The method may include processing data of the subject using a natural language processing algorithm, an audio processing algorithm, or a video processing algorithm to identify or assess the mental state of the subject, where the data includes speech or video data of the subject, and the method may include outputting a report indicative of the mental state of the subject. The report may be transmitted to a user for use in identifying, assessing, or monitoring the mental state.

[0070] In some embodiments, the user may be the subject. In some embodiments, the user may be a clinician or medical professional. In some embodiments, the report may include a plurality of graphic visual elements. In some embodiments, the report may be configured to be displayed on a graphical user interface of the user's electronic device. In some embodiments, the method may further include updating the report in response to one or more detected changes in mental state of the subject. In some embodiments, the report may be updated in near real-time as one or more detected changes in mental state occur in the subject.

[0071] Another aspect of the present disclosure provides a method for identifying whether a subject is at risk for a mental or physiological condition. The method may include obtaining speech data from the subject and storing the speech data in a computer memory, processing the speech data, in part, using natural language processing to identify one or more features indicative of the mental or physiological condition, and outputting an electronic report identifying whether the subject is at risk for a mental or physiological condition, where the risk may be quantified in the form of a normalized score with a confidence level. The normalized score with the confidence level may be usable by a user to identify whether the subject is at risk for a mental or physiological condition.

[0072] In some embodiments, the user may be a subject. In some embodiments, the user may be a clinician or medical professional. In some embodiments, the report may include a number of graphic visual elements. In some embodiments, the report may be configured to be displayed on a graphical user interface of the user's electronic device.

[0073] Another aspect of the present disclosure provides a method for identifying, assessing, or monitoring a mental condition or disorder of a subject, the method may include receiving audio or audiovisual data including speech of the subject in a computer memory, processing the audio or audiovisual data to identify, assess, monitor, or diagnose a mental condition or disorder of the subject, the processing may include performing natural language processing on the speech of the subject.

[0074] In some embodiments, the audio or audiovisual data may be received in response to a query directed to the subject. In some embodiments, the audio or audiovisual data may be from a pre-recording of a conversation to which the subject may be a party. In some embodiments, the audio or audiovisual data may be from a pre-recording of a clinical session involving the subject and a medical professional. In some embodiments, a mental condition or disorder may be identified with a high performance level compared to a standardized mental health questionnaire or testing tool. In some embodiments, the processing may further include performing an acoustic analysis on the subject's speech using the trained algorithm.

[0075] Another aspect of the present disclosure provides a method for inferring whether a subject has a mental condition and providing the inference to an interested party. The method can include obtaining speech data from the subject and storing the speech data in a computer memory. The speech data can include responses to a plurality of queries transmitted to the subject in audio and / or visual form. The method can include selecting (1) a first model optimized for sensitivity in inferring whether the subject has the mental condition or (2) a second model optimized for specificity in inferring whether the subject has the mental condition. The method can further include processing the speech data using the selected first model or second model to generate an inference. The method can further include transmitting the inference to the interested party.

[0076] In some embodiments, the first model can be selected and the stakeholder can be a healthcare payer. In some embodiments, the second model can be selected and the stakeholder can be a healthcare provider.

[0077] Another aspect of the present disclosure provides a system for determining whether a subject is at risk for having a mental condition. The system may be configured to (i) receive speech data from a memory, and (ii) process the speech data using at least one model to determine that the subject is at risk for having a mental condition. The at least one model may be trained on speech data from other subjects who have clinical determinations of the mental condition. The clinical determinations may serve as indicators for the speech data. The system may be configured to generate an estimate of the mental condition that is in accordance with one or more performance metrics compared to a clinically validated survey, test, or questionnaire.

[0078] In some embodiments, the system may be configured to generate an estimate of the mental state with high specificity compared to clinically validated surveys, tests, or questionnaires. In some embodiments, the system may be configured to generate an estimate of the mental state with high sensitivity compared to clinically validated surveys, tests, or questionnaires. In some embodiments, the indicia may be output while the subject is speaking. In some embodiments, the indicia may be output by a streaming or periodically updated signal.

[0079] Another aspect of the present disclosure provides a method for assessing a mental state of a subject. The method may include using an automated screening module to dynamically construct at least one query based in part on one or more target mental states to be assessed. The at least one query may be configured to elicit at least one response from the subject. The method may further include transmitting the at least one query in audio and / or visual form to the subject to elicit at least one response. The method may further include receiving data including at least one response from the subject in response to transmitting the at least one query. The data may include speech data. The method may further include processing the data using a composite model including at least one or more semantic models to generate an assessment of the physical state of the subject.

[0080] Another aspect of the present disclosure provides a non-transitory computer-readable medium that includes machine-executable instructions that, when executed by one or more computer processors, perform any of the methods described above or elsewhere herein.

[0081] Another aspect of the present disclosure provides a system comprising one or more computer processors and a computer memory coupled thereto, the computer memory including machine-executable code that, when executed by the one or more computer processors, performs any of the methods above or elsewhere herein.

[0082] Further aspects and advantages of the present disclosure will become readily apparent to those skilled in the art from the following detailed description, in which merely exemplary embodiments of the present disclosure are shown and described. As will be understood, the present disclosure is capable of other and different embodiments, and its several details are capable of modification in various obvious respects, all without departing from the present disclosure. Accordingly, the drawings and description should be regarded as illustrative in nature, and not as restrictive. INCORPORATION BY REFERENCE All publications, patents, and patent applications mentioned in this specification are herein incorporated by reference to the same extent as if each individual publication, patent, or patent application was specifically and individually indicated to be incorporated by reference. To the extent that the publications and patents or patent applications incorporated by reference conflict with a disclosure contained in the specification, the specification is intended to take precedence and / or supersede any such conflicting material.

[0083] The novel features of the present disclosure are set forth with particularity in the appended claims. A better understanding of the features and advantages of the present disclosure can be obtained by reference to the following detailed description that sets forth illustrative embodiments, in which the principles of the disclosure are utilized, and the accompanying drawings (also referred to herein as "FIGS." and "FIGS."), in which: [Brief description of the drawings]

[0084] [Figure 1A] Health Screening or Monitoring Server and Clinical Data Server A health screening or monitoring system is shown in which a computer system and a social data server cooperate to estimate a patient's health status in accordance with the present disclosure. [Figure 1B] 1B shows a further embodiment of the health screening or monitoring system from FIG. 1A. [Diagram 2] 1 illustrates a patient screening or monitoring system according to some embodiments in which a web server and modeling server(s) cooperate to assess a patient's condition over a wide area network. [Diagram 3] A patient evaluation system according to the present disclosure is shown in which a real-time computer system, a modeling computer system, and a clinical and demographic data server computer system work together to evaluate a patient's condition and report the evaluated condition to a clinician using a clinician device over a wide area network. [Figure 4] FIG. 1B is a more detailed block diagram of the health screening or monitoring server of FIG. 1A. [Diagram 5] FIG. 5 is a more detailed block diagram of the interactive health screening or monitoring logic of the health screening or monitoring server of FIG. 4. [Figure 6] FIG. 6 is a more detailed block diagram of the interactive screening or monitoring server logic of the interactive health screening or monitoring logic of FIG. 5. [Figure 7] FIG. 7 is a more detailed block diagram of the generalized dialogue flow logic of the interactive screening or monitoring server logic of FIG. [Figure 8] FIG. 1 is a logic flow diagram illustrating control of an interactive verbal conversation with a patient by a generalized dialogue flow logic according to the present disclosure. [Figure 9] FIG. 7 is a more detailed block diagram of the Generic Dialogue Flow Logic Question-Adaptive Action Bank. [Figure 10] FIG. 9 is a more detailed logic flow diagram of the steps of FIG. 8. [Figure 11] FIG. 10 is a more detailed block diagram of the question management logic of the Question-Adaptive Action Bank of FIG. [Figure 12] FIG. 13 is a logic flow diagram for determining the quality of a question according to the present disclosure. [Figure 13] FIG. 13 is a logic flow diagram for determining equivalence of two questions according to the present disclosure. [Figure 14] FIG. 1 is a logic flow diagram illustrating control of an interactive verbal conversation with a patient by a real-time system according to the present disclosure. [Figure 15] FIG. 15 is a more detailed logic flow diagram of each step of FIG. 14. [Figure 16]FIG. 15 is a more detailed logic flow diagram of each step of FIG. 14. [Figure 17] FIG. 4 is a transaction flow diagram showing an illustrative example of a verbal conversation with and controlled by the real-time system of FIG. [Figure 18] FIG. 4 is a more detailed block diagram of the runtime model server logic of the interactive health screening or monitoring logic of FIG. 3. [Figure 19] FIG. 1B is a more detailed block diagram of the model training logic of the interactive health screening or monitoring logic of FIG. 1A. [Figure 20A] FIG. 1 shows a more detailed block diagram of a patient screening or monitoring system according to some embodiments. [Figure 20B] 1 provides a block diagram of a runtime model server(s) according to some embodiments. [Figure 21] 1 provides a block diagram of a model training server(s) according to some embodiments. [Figure 22] 4 illustrates in more detail the real-time computer system and the modeling computer system of FIG. 3, including the general flow of data. [Figure 23A] 1 provides a block diagram of an acoustic model according to some embodiments. [Figure 23B] 23B shows an embodiment of FIG. 23A including an acoustic modeling block. [Figure 23C] The score calibration and confidence module is shown. [Figure 24] For illustrative purposes, we provide a simplified example of a high-level feature representator for an acoustic model. [Diagram 25] 1 provides a block diagram of a natural language processing (NLP) model according to some embodiments. [Figure 26] 1 provides a block diagram of a visual model according to some embodiments. [Figure 27] 1 provides a block diagram of descriptive features according to some embodiments. [Figure 28]1 provides a block diagram of a dialogue engine according to some embodiments. [Figure 29] FIG. 1 is a logic flow diagram of an example process for screening a patient for a mental health condition according to some embodiments. [Diagram 30] FIG. 1 is a logic flow diagram of an example process for model training according to some embodiments. [Diagram 31] FIG. 1 is a logic flow diagram of an example process for model personalization according to some embodiments. [Diagram 32] FIG. 2 is a logic flow diagram of an example process for client interaction according to some embodiments. [Diagram 33] FIG. 1 is a logic flow diagram of an example process for classifying a mental state of a client according to some embodiments. [Diagram 34] FIG. 1 is a logic flow diagram of an example process for model adjustment according to some embodiments. [Diagram 35] FIG. 13 is a logic flow diagram of an example process for model weighting and fusion according to some embodiments. [Diagram 36] FIG. 2 is a logic flow diagram of a simplified example process for acoustic analysis, provided for illustrative purposes only. [Figure 37] FIG. 2 is a block diagram showing the speech recognition logic of the modeling computer system in more detail. [Figure 38] FIG. 2 is a block diagram illustrating the language model training logic of the modeling computer system in more detail. [Figure 39] FIG. 2 is a block diagram showing the language model logic of the modeling computer system in more detail. [Diagram 40] FIG. 2 is a block diagram showing the acoustic model training logic of the modeling computer system in greater detail. [Diagram 41] FIG. 2 is a block diagram showing the acoustic model logic of the modeling computer system in more detail. [Diagram 42] FIG. 2 is a block diagram showing the visual model training logic of the modeling computer system in greater detail. [Diagram 43] FIG. 2 is a block diagram showing the visual model logic of the modeling computer system in more detail. [Diagram 44] FIG. 1B is a more detailed block diagram of the screening or monitoring system data store of the interactive health screening or monitoring logic of FIG. 1A. [Diagram 45] 1 illustrates a health screening or monitoring system in which a health screening or monitoring server estimates a patient's health status by passively listening to surrounding speech in accordance with the present disclosure. [Diagram 46] FIG. 1 is a logic flow diagram illustrating estimating a patient's health status by passively listening to ambient speech in accordance with the present disclosure. [Figure 47] FIG. 1 is a logic flow diagram illustrating estimating a patient's health status by passively listening to ambient speech in accordance with the present disclosure. [Figure 48] FIG. 5 is a more detailed block diagram of the health care management logic of the health screening or monitoring server of FIG. 4. [Figure 49] FIG. 49 is a component state block diagram of the healthcare management logic workflow of FIG. 48. [Figure 50] FIG. 49 is a block diagram of the actions of the healthcare management logic workflow of FIG. 48. [Figure 51] FIG. 49 is a logic flow diagram for automated construction of the healthcare management logic workflow of FIG. 48 in accordance with the present disclosure. [Figure 52] FIG. 4 is a more detailed block diagram of the real-time computer system of FIG. [Diagram 53] FIG. 4 is a more detailed block diagram of the modeling computer system of FIG. 3. [Figure 54] FIG. 1B is a more detailed block diagram of the health screening or monitoring server of FIG. 1A. [Figure 55] 1 provides an illustrative example of a spectrogram of an acoustic signal used for analysis according to some embodiments. [Figure 56]1 provides an illustrative example of a spectrogram of an acoustic signal used for analysis according to some embodiments. [Figure 57] 1 is an illustrative example of a computer system in which the present disclosure may be implemented. [Figure 58] 1 is an illustrative example of a computer system in which the present disclosure may be implemented. [Figure 59] Demonstrate precision case management use cases for the system. [Figure 60] 1 illustrates a primary care screening or monitoring use case for the system. [Figure 61] A system for advanced employee assistance plan (EAP) navigation and triage is presented. [Figure 62] 1 illustrates a computer system that is programmed or otherwise configured to assess the mental state of a subject in a single session or across multiple different sessions. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0085] While various embodiments of the present invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are given by way of example only. Many variations, changes, and substitutions may occur to those skilled in the art without departing from the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be used.

[0086] Aspects, features, and advantages of exemplary embodiments of the present invention will become better understood with respect to the following description in conjunction with the accompanying drawing(s). As will be appreciated by those skilled in the art, the described embodiments of the present invention provided herein are merely illustrative and not limiting, and are provided by way of example only. All features disclosed in this specification body may be replaced by alternative features serving the same or similar purpose, unless otherwise specified. Thus, many other embodiments of its modifications are considered to be within the scope of the present invention as defined herein and its equivalents.

[0087] The use of absolute and / or sequential terms hereinafter, e.g., "shall", "shall not", "shall not", "shall", "must", "must not", "firstly", "firstly", "next", "then", "before", "after", "lastly", and "finally", is not intended to limit the scope of the invention, as the embodiments disclosed herein are merely exemplary. The present invention relates to health screening or monitoring systems, and more particularly to computer-implemented mental health screening or monitoring tools that utilize language analysis, visual cues, and acoustic analysis to greatly improve accuracy and effectiveness. In this application, details of improved acoustic, visual, and speech analysis techniques are described as they relate to classifying respondents as depressed or other mental conditions of interest. While most of the disclosure below focuses primarily on assessing depression in patients, the systems and methods described herein can successfully screen or monitor users for a myriad of mental and physical ailments. For example, bipolar disorder, anxiety, and schizophrenia are examples of mental disorders that such systems can successfully screen for or monitor. It is also envisioned that such systems can be utilized to assess physical ailments. While this disclosure may focus on depression screening or monitoring, it should be understood that this is not limiting: any suitable mental or physical illness may be screened for using the disclosed systems and methods.

[0088] The systems and methods disclosed herein may perform semantic analysis on patient utterances using natural language processing (NLP). Semantic analysis as disclosed herein may refer to analyzing spoken language from patient responses to assessment questions or captured conversations to determine the meaning of the spoken language for the purpose of performing mental health screening or monitoring of the patient. The analysis may be of words or phrases and may be configured to consider primary or follow-up queries. In the case of captured human-to-human conversations, the analysis may be applied to the other person's utterances. As used herein, the terms "semantic analysis" and "natural language processing (NLP)" may be used interchangeably. Semantic analysis may be used to determine the meaning of the patient's utterance in context. Semantic analysis may also be used to determine the topic the patient is talking about.

[0089] Mental states as described herein may be distinguished from emotions or feelings such as happiness, sadness, or anger. Mental states may include one or more emotions combined with a philosophy of mind, including how a person perceives objects in their environment and the behavior of others toward them. Although emotions may be temporary, mental states may represent a person's overall disposition or mood, even in situations where the person's emotions may change. For example, a person who is depressed may feel variously happy, sad, or angry at different times.

[0090] According to one or more embodiments of the present invention, a server computer system (health screening or monitoring server 102 - FIG. 1A) may administer health condition screening or monitoring tests to a human patient using a client device (patient device 112) by engaging the patient in an interactive verbal conversation and applying a composite model, which may combine language models, acoustic models, metadata models, and visual models, to the captured audiovisual signals of the patient involved in the conversation. The general subject matter of the screening or monitoring test may be similar to that of a standardized depression screening or monitoring test such as the PHQ-9, but the composite model may analyze the patient's audiovisual signals in real time to (i) make the conversation more appealing to the patient, and (ii) estimate the patient's health condition. Appendix A illustrates exemplary implementations including a scheduling service, an SMS service, a dialogue service, a call service, and a user management service. While the latter goal is primary, the former goal is a key factor in achieving the latter. The patient's truthfulness in answering questions posed by the screening or monitoring test is important in assessing the patient's mood. The health screening or monitoring server 102 encourages patient integrity.

[0091] First, the verbal conversation may give the patient little time to make an insincere response to the question, instead of simply responding to the question honestly. Second, the conversation may feel more spontaneous and personal to the patient, and may be less intrusive to the patient, than a general questionnaire, such as that provided by simply administering the PHQ-9. Thus, the verbal conversation may not induce or amplify the patient's resentment for having to answer a questionnaire for the benefit of the physician or other clinician. Third, the verbal conversation may be adapted on the fly to respond to the patient, thereby reducing the patient's intrusiveness to the screening or monitoring test, and in some cases, the screening or monitoring test may be shortened. Fourth, the screening or monitoring test performed by the health screening or monitoring server 102 may further rely on non-verbal aspects of the conversation in addition to the verbal content of the conversation to evaluate the patient's depression. As shown in FIG. 1A, the health screening or monitoring system 100 may include a health screening or monitoring server 102, a call center system 104, a clinical data server 106, a social data server 108, a patient device 112, and a clinician device 114, which are connected to each other via a wide area network (WAN) 110, which in this exemplary embodiment is the Internet. In this exemplary embodiment, the patient device 112 may be contactable by the call center system 104 via a public switched telephone network (PSTN) 120 or directly. The health screening or monitoring server 102 may be a server computer system that administers health screening or monitoring tests to patients via the patient devices 112 and combines language models, acoustic models, and visual models to generate results 1820 (FIG. 18), using clinical data retrieved from the clinical data server 106, social data retrieved from the social data server 108, and patient data collected from past screenings or monitoring to train models in the runtime model server 304 (FIG. 18).The clinical data server 106 (FIG. 1A) may be a server computer system that makes available a patient's available clinical or demographic data, including diagnosis, medication information, and the like, to, for example, the health screening or monitoring server 102, in a manner that complies with HIPAA (Health Insurance Portability and Accountability Act of 1996) and / or any other privacy and security policies and regulations, such as GDPR and SOC2. The social data server 106 may be a server computer system that makes available a patient's social data, including social media posts, online purchases, searches, and the like, to, for example, the health screening or monitoring server 102. The clinician device 114 may be a client device that receives data representing the results of screening or monitoring related to the patient's health from the health screening or monitoring server 102.

[0092] The system may be used to assess the mental state of a subject in a single session or across multiple sessions. Subsequent sessions may be informed by assessment results from previous assessments. This may be done by providing assessment data as input to machine learning algorithms or other analytical methods for subsequent assessments. Each session may generate one or more assessments. An individual assessment may compile data from multiple sessions.

[0093] Figure 1B illustrates a further embodiment of the health screening or monitoring system from Figure 1A. Figure 1B illustrates a conversation between a patient 120 and a clinician 130. The clinician 130 may record one or more speech samples from the patient 120 and, with the patient's 120 consent, upload those samples to the wide area network 110. The speech samples may be analyzed by one or more machine learning algorithms described elsewhere herein.

[0094] 2 provides a further embodiment of a health screening or monitoring system. The health screening or monitoring system 200 may apply health condition screening or monitoring tests to a human patient using a client device (clients 260a-n) by engaging the patient in a dialogue and applying a composite model that combines a language model, an acoustic model, and a visual model to the captured audiovisual signals of the patient involved in the dialogue.

[0095] While the general subject matter of the screening or monitoring test may be similar to that of a standardized depression screening or monitoring test such as the PHQ-9, the combined model may be configured to analyze the patient's audiovisual signals in real time to (i) make the conversation more engaging for the patient, (ii) estimate the patient's mental health, and (iii) provide a judgment-free and less embarrassing experience for patients who may already suffer from anxiety or other mental barriers to receiving appropriate screening or monitoring from a clinician.

[0096] It should be noted that throughout this disclosure, a set of terms may be used interchangeably, and this usage is not intended to limit the scope of the disclosure in any manner. For example, the terms "patient," "client," "subject," "respondent," and "user" may all be used interchangeably to refer to individuals being screened for mental health conditions and / or devices utilized by the individuals to collect and transmit audio and visual data used to screen them. Similarly, "semantic analysis" and "NLP" may be used interchangeably to refer to natural language processing models and elements. Similarly, "stakeholders" are used to refer to a wide variety of interested third parties that are not the patient being screened. These stakeholders may include physicians, medical personnel, care team members, insurance companies, research institutions, and the patient's family / relatives, hospitals, crisis centers, etc. Thus, when other labels such as "physician" are used, it should be understood that the intent of the disclosure is to refer to any number of stakeholders.

[0097] The health screening or monitoring system 200 includes a back-end infrastructure designed to perform screening or monitoring interactions and analyze results. It includes one or more model servers 230 coupled to a web server 240. The web server 240 and the model server(s) 230 utilize user data 220 that is further populated by clinical and social data 210. The clinical data portion may be compiled from medical personnel and may include diagnoses, vital information (age, weight, height, blood chemistry, etc.), illnesses, medications, clinical encounter listings (hospitalizations, clinic visits, emergency department visits), clinician notes, etc. This clinical data may be compiled from one or more electronic health record (EHR) systems or health information exchanges (HIE) via secure application protocols, extensions, or sockets. The social data may include information collected from the patient's social network including social media posts, from databases detailing the patient's purchases, and from databases including the patient's economic, educational, residential, legal, and other social determinants. This information may be compiled along with further preference data, metadata, annotations, and voluntarily provided information to populate the user database 220. The model server 230 and web server 240 may further populate and / or augment the user data 220 with preferences, extracted features, and the like.

[0098] The backend infrastructure communicates with clients 260a-n through network infrastructure 250. Typically, this network includes the Internet, a corporate local area network, a private intranet, a cellular network, or any combination of these. Clients 260a-n include the client devices of the person being screened, which access the backend screening or monitoring system and include a microphone and camera for audio and video capture. The client devices may be mobile phones, tablets, laptops or desktops with microphones and optional cameras, smart speakers in the home or elsewhere, smart watches with microphones and optional cameras, or similar devices.

[0099] Client devices may collect additional data, such as biometric data. For example, smart watches and fitness trackers already have the ability to measure movement, heart rate, and possibly respiration rate and blood oxygenation, as well as other physiological parameters. Future smart devices may record conductivity measurements to track sweat, skin pH changes, and other chemical or hormonal changes. Client devices may work in concert to collect data. For example, a phone call may capture audio and visual data, while a Bluetooth-enabled fitness tracker may simultaneously provide temperature, pulse rate, respiration rate, and movement data.

[0100] All data collected for each client 260a-n is provided to the originating web server 240 via the network infrastructure 250. After processing, the results are provided to the originating client 260a-n for consumption and, if desired, sharing with one or more stakeholders 270a-n, respectively, associated with a given client 260a-n. In this illustrative example, the stakeholders 270a-n are shown to be in direct communication with their respective clients 260a-n. While in practice this may indeed be envisaged, in many cases the stakeholders 270a-n may have direct access to a back-end screening or monitoring system via the network infrastructure 250 and the web server 240, without the need to use the client 260a-n as an intermediary. However, FIG. 2 provides this configuration to more clearly illustrate that each client 260a-n may be associated with one or more stakeholders 270a-n, which may be different from the stakeholders 270a-n of any other client 260a-n.

[0101] In another embodiment of the screening or monitoring system, a server computer system (real-time system 302 - Figs. 3 and 22) applies a depression assessment test to a human patient using a client device (portable device 312) by engaging the patient in an interactive verbal conversation and applying a composite model that combines language, acoustic, metadata, and visual models to the captured audiovisual signals of the patient involved in the conversation. The assessment does not only include analysis of responses to survey questions, although the general subject matter of the assessment test may incorporate queries that include similar subjects to questions asked in standardized depression assessment tests such as the PHQ-9. In effect, the composite model of the screening or monitoring system analyzes the patient's audiovisual signals in real time to (i) make the conversation more engaging for the patient, and (ii) assess the patient's mental health.

[0102] While the latter goal is the primary goal, the former goal is an important factor in achieving the latter. The patient's truthfulness in answering questions posed by the assessment test is important in assessing the patient's mood. The real-time system 302 encourages patient truthfulness in several ways. First, verbal conversation allows the patient little time to formulate a response to a question, rather than simply committing to responding to the question truthfully.

[0103] Second, the conversation feels more spontaneous and personal to the patient than a clearly general questionnaire, and is less intrusive to the patient. Thus, the verbal conversation does not induce or increase the patient's resentment for having to answer a questionnaire before seeing a doctor or other clinician. Third, the verbal conversation is adapted on the fly to respond to the patient, thereby reducing the patient's intrusiveness to the assessment test and, in some cases, shortening the assessment test. Fourth, the assessment test performed by the real-time system 302 may rely on the conversation and the patient's non-verbal aspects rather than the verbal content of the conversation to assess the patient's depression.

[0104] As shown in (FIG. 3), the patient evaluation system 300 includes a real-time system 302, a modeling system 304, a clinical data server 306, a patient device 312, and a clinician device 314, which are connected to each other via a wide area network (WAN) 310, which in this exemplary embodiment is the Internet. The real-time system 302 is a server computer system that administers depression evaluation tests to patients via the patient device 312. The modeling system 304 is a server computer system that combines a number of language models, acoustic models, and visual models to generate a composite model 2204 (FIG. 22), and uses clinical data retrieved from the clinical data server 306 and patient data collected from past evaluations to train the composite model 2204. The clinical data server 306 (FIG. 3) is a server computer system that makes available to the monitoring system 304, for example, patient clinical data including diagnosis, medication information, etc., in a manner that complies with HIPAA (Health Insurance Portability and Accountability Act of 1996) and / or any other privacy and security policies and regulations, such as GDPR and SOC2. The clinician device 314 is a client device that receives data representing the resulting assessment regarding depression from the real-time system 302 .

[0105] The end-to-end nature of the system The system disclosed herein may provide the medical professional with a prediction of the patient's mental state. The mental state may be depression, anxiety, or other mental state. The system may provide additional information to the medical professional beyond the mental state prediction. The system may provide demographic information, such as age, weight, occupation, height, ethnicity, medical history, psychological history, and gender, to the medical professional via a client device, such as client devices 260a-n of FIG. 2. The system may provide the information from an online system or social network where the patient may be registered. The patient may opt-in by setting the permission to provide this information on their client device before the screening or monitoring process begins. The patient may also be prompted to enter demographic information during the screening or monitoring process. The patient may choose to provide information to the medical professional from their electronic health record. Additionally, the medical professional may interview the patient during or after the screening or monitoring event to obtain demographic information. During screening or monitoring registration, the patient may enter information identifying or limiting their interests. For example, the patient may enter topics they would like to talk about or would not like to talk about. In this disclosure, the terms "healthcare practitioner" and "clinician" are used interchangeably. A healthcare practitioner may be a doctor, nurse, physician assistant, nursing assistant, clinical psychologist, social worker, technician, or other healthcare provider.

[0106] The clinician may set up a mental health assessment for the patient. This may include selecting a list of questions that the system will ask the patient, including follow-up questions. The clinician may add or remove specific questions from the assessment or change the order in which the questions are administered to the patient. The clinician may be available as a proctor during the assessment to answer any clarifying questions the patient may have.

[0107] The system may provide the clinician with a dialogue between itself and the patient. This dialogue may be a recording of a screening or monitoring process or a textual transcription of the dialogue. The system may provide a summary of the dialogue between itself and the patient using semantic analysis to select segments of speech that were most significant for predicting the patient's mental state. These segments may be selected, for example, because they may be highly weighted in the calculation of a binary or scaled score indicative of mental state prediction. The system may incorporate such generated scores into a summary report about the patient, along with semantic context obtained from a transcription of the interview with the patient.

[0108] The system may further provide the clinician with a "word cloud" or "topic cloud" extracted from the text transcription of the patient's speech. A word cloud may be a visual display of individual words or phrases, where the most frequently used words and phrases are designated using a larger font size, different colors, different fonts, different typefaces, or any combination of these. Because depressed patients generally say certain words or phrases more frequently than non-depressed patients, depicting the frequency of words or phrases in such a manner may be beneficial. For example, depressed patients may use words or phrases that exhibit dark, black, or morbid humor. They may talk about feeling worthless or like they have failed, or use absolutist language such as "always," "never," and "completely." Depressed patients may also use first-person singular pronouns (e.g., "I," "me") more frequently and second- or third-person pronouns less frequently compared to the general population. The system may train a machine learning algorithm to perform semantic analysis of word clouds of depressed and non-depressed people so that people can be classified as depressed or non-depressed based on those word clouds. The word cloud analysis may be performed using unsupervised learning. For example, the system may analyze unlabeled word clouds to search for patterns to separate people into groups based on their mental state.

[0109] The systems described herein can output an electronic report identifying whether a patient is at risk for a mental or physiological condition. The electronic report can be configured to be displayed on a graphical user interface of the user's electronic device. The electronic report can include a quantification of the risk for the mental or physiological condition, e.g., a normalized score. The score can be normalized for the entire population of subjects or for a subpopulation. The electronic report can also include a confidence level for the normalized score. The confidence level can indicate the reliability of the normalized score (i.e., the degree to which the normalized score can be trusted).

[0110] The electronic report may include a visual graphical element, for example, if a patient has multiple scores from multiple screening or monitoring sessions that occurred at several different times, the visual graphical element may be a graph showing the progression of the patient's scores over time.

[0111] The electronic report may be output to the patient or a contact associated with the patient, a medical professional, a medical payer, or other third party. The electronic report may be output in substantially real-time, even while the screening, monitoring, or diagnosis is ongoing. In response to changes in the normalized score or confidence level during the course of the screening, monitoring, or diagnosis, the electronic report may be updated and resent to the user in substantially real-time.

[0112] In some cases, the electronic report may include one or more descriptors regarding the patient's mental state. The descriptors may be qualitative indicators of the patient's mental state (e.g., "mild depression"). Alternatively or additionally, the descriptors may be topics mentioned by the patient during screening. The descriptors may be displayed graphically, for example, in a word cloud.

[0113] The models described herein may be optimized for a particular purpose or based on the entity that may receive the output of the system. For example, the models may be optimized for sensitivity in predicting whether a patient has a mental condition. Healthcare payers, such as insurance companies, may prefer such models so that they can minimize the number of claims made for patients with false positive diagnoses. In other cases, the models may be optimized for specificity in predicting whether a patient has a mental condition. Healthcare professionals may prefer such models. The system may select an appropriate model based on the stakeholder to which the output is sent. After processing, the system can send the output to the stakeholder.

[0114] Alternatively, the models described herein may be tuned or configured to process speech and other data according to a desired level of sensitivity or a desired level of specificity as determined by a clinician, medical professional, insurance company, or government regulatory agency. Use Cases

[0115] The system may be used to monitor depression in teenagers. The system may perform machine learning analysis on a group of teenagers to determine voice-based biomarkers that can uniquely classify teenagers as at risk for depression. Depression in teenagers may have different causes than in adults. Hormonal changes may result in teenagers' behaviors that are atypical for adults. Systems for screening or monitoring teenagers need to employ models that are tailored to recognize these unique behaviors. For example, a depressed or upset teenager may be more prone to becoming angry or irritable than an adult who may withdraw when upset. Thus, questions from the assessment may elicit different voice-based biomarkers from teenagers than adults. Different screening or monitoring methods than those used to screen or monitor adults may be used when testing teenagers for depression or when examining the mental state of teenagers. Clinicians may modify the assessment to elicit voice-based biomarkers that are specific to depression in teenagers in particular. The system may be trained using these assessments to determine a teen-specific model for predicting mental states. Teens may be further segmented by home (foster care, adoptive parent(s), two biological parents, one biological parent, guardian / relative care, etc.), medical history, gender, age (elderly vs. younger teens), and socio-economic status, and these segments may be incorporated into the model's predictions.

[0116] The system may be used to monitor the elderly for depression and dementia. Elderly people may also have certain voice-based biomarkers that younger adults may not have. For example, elderly people may have a strained or thin voice due to aging. Elderly people may exhibit aphasia or dysarthria, struggle to understand survey questions, follow-ups, or conversational utterances, and may use repetitive words. Clinicians may develop surveys to elicit certain voice-based biomarkers from elderly patients, or algorithms may be used to develop such surveys. Machine learning algorithms may be developed to predict the mental state of elderly patients, specifically by segmenting patients by age. Differences may exist between elderly patients and different generations (e.g., Greatest, Silent, Boomers), who may have different views on gender roles, morality, and cultural norms. Models may be trained to incorporate the age group, gender, race, socioeconomic status, physical medical conditions, and family involvement of the elderly.

[0117] The system may be used to test the mental fitness of airline pilots, who have a very demanding job and may experience large amounts of stress and fatigue on long-haul flights. Clinicians or algorithms may be used to develop screening or monitoring methods for these conditions. For example, the system may base the assessment from queries similar to those tested in the Minnesota Multiphasic Personality Inventory (MMPI) and MMPI-2.

[0118] The system may be used to screen military personnel for mental fitness. For example, the system may conduct an assessment to test for PTSD using queries with themes similar to those asked in the Diagnostic and Statistical Manual of Mental Disorders (DSM)-5 for Primary Care Posttraumatic Stress Disorder (PC-PTSD-5). In addition to PTSD, the system may screen military personnel for depression, panic disorder, phobic disorder, anxiety, and hostility. The system may employ different surveys to screen military personnel pre-deployment and post-deployment. The system may segment military personnel by segmenting on occupation, as well as by branch, commissioned or enlisted, gender, age, ethnicity, length of service / number of deployments, marital status, medical history, and other factors.

[0119] The system may be used to evaluate a prospective gun purchaser, such as by conducting a background check. The evaluation may be designed by a clinician or algorithmically to evaluate the prospective purchaser for mental fitness to own a firearm. The survey may have a requirement to determine, using questions and follow-up questions, whether the prospective gun purchaser may be found by a court or other authority to be a danger to himself or herself or others.

[0120] The health screening or monitoring server 102 (FIG. 1A) is shown in further detail in FIG 4 and in even greater detail in FIG 22. As shown in FIG 4, the health screening or monitoring server 102 includes interactive health screening or monitoring logic 402 and health care management logic 408. Additionally, the health screening or monitoring server 102 includes a screening or monitoring system data store 410 and a model repository 416.

[0121] Each component of the health screening or monitoring server 102 is described more fully herein. Briefly, the interactive health screening or monitoring logic 402 engages in an interactive dialogue with a subject patient to estimate one or more health conditions of the patient by applying models from the runtime model server 504 (FIG. 18) to audiovisual signals representative of responses by the patient. In this exemplary embodiment, the interactive health screening or monitoring logic 402 (FIG. 4) may operate in a passive listening mode to observe the patient outside the context of an interactive dialogue with the health screening or monitoring server 102, such as during a session with a medical clinician, and estimate the patient's health condition from such observations. The healthcare management logic 408 makes expert recommendations in response to the health condition estimations of the interactive health screening or monitoring logic 402. The screening system data store 410 stores and maintains all user and patient data required for and collected by screening or monitoring in the manner described herein.

[0122] The conversational context of a health screening or monitoring system may improve one or more performance metrics associated with one or more machine learning algorithms used by the system. These metrics may include metrics such as F1 score, area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), and equal error rate.

[0123] It should be understood that the behavior of the health screening or monitoring server 102 described herein may be distributed across multiple computer systems. For example, in some exemplary embodiments, the real-time interactive behavior of the health screening or monitoring server 102 (e.g., interactive screening or monitoring server logic 502 and runtime model server logic 504 described below) is implemented on one or more servers configured to handle large amounts of traffic traversing the WAN 110 (FIG. 1A), and the computationally intensive behavior of the health screening or monitoring server 102 (e.g., healthcare management logic 408 and model training logic 506) is implemented on one or more other servers configured to efficiently perform highly complex calculations. The distribution of the various loads carried by the health screening or monitoring server 102 may be distributed among multiple computer systems.

[0124] The interactive health screening or monitoring logic 402 is shown in more detail in Figure 5. The interactive health screening or monitoring logic 402 includes interactive screening or monitoring server logic 502, runtime model server logic 504, and model training logic 506. The interactive screening or monitoring server logic 502 conducts the interactive screening or monitoring conversation with a human patient, the runtime model server logic 504 uses and trains several machine learning models to simultaneously evaluate the patient's responsive audiovisual signals, and the model training logic 506 trains the models of the runtime model server logic 504.

[0125] The interactive screening or monitoring server logic 502 is shown in more detail in Figure 6 and includes generic dialogue flow logic 602 and input / output (I / O) logic 604. The I / O logic 604 affects the interactive screening or monitoring conversation by sending audiovisual signals to and receiving audiovisual signals from the patient device 112. The I / O logic 604 receives data from the generic dialogue flow logic 602 that specifies questions to be asked of the patient and sends audiovisual data representing those questions to the patient device 112. In embodiments in which the interactive screening or monitoring conversation is conducted over the PSTN 120 (FIG. 1), the I / O logic 604 (i) sends audiovisual signals to the patient device 112 by sending data to a human or automated operator at the call center 104 to prompt the operator to ask a question over the phone with the patient device 112 (or alternatively, by sending data to a back-end automated dialogue system for the patient), and (ii) receives audiovisual signals from the patient device 112 by receiving audiovisual signals of the interactive screening or monitoring conversation forwarded by the call center 104. The I / O logic 604 also sends at least a portion of the received audiovisual signals of the interactive screening or monitoring conversation to the runtime model server logic 504 (FIG. 18) and the model training logic 506 (FIG. 19).

[0126] The queries or questions asked to the patient may be stored as nodes, while the patient responses collected as audiovisual signals may be stored as edges. Thus, a screening or monitoring event or a set of screening or monitoring events for a particular patient may be represented as a graph. For example, different answers to different follow-up questions may be represented as multiple spokes connecting a particular node to multiple other nodes. Different graph structures for different patients may be used as training examples for machine learning algorithms as another way to determine mental state classification for a patient. Classification may be performed, for example, by determining the similarity between graphs of depressed patients. As discussed herein, equivalent questions may be labeled in this way in the graph. Thus, the graph may be studied and analyzed to determine specificity in the interpretation of different versions of questions by patients.

[0127] The I / O logic 604 also receives results 1820 (FIG. 18) from the runtime server logic 504 that represent an evaluation of the audiovisual signals. The generic dialogue flow logic 602 conducts an interactive screening or monitoring conversation with a human patient. The generic dialogue flow logic 602 determines what questions the I / O logic 604 should ask the patient and monitors the patient's responses, which are represented in results 1820. Additionally, the generic dialogue flow logic 602 determines when to gracefully end the interactive screening or monitoring conversation.

[0128] The generic dialogue flow logic 602 is shown in more detail in Figure 5. The generic dialogue flow logic 602 includes a dialogue control logic generator 702. The dialogue control logic generator 702 manages an interactive screening or monitoring conversation with a patient by sending data representing dialogue actions to the I / O logic 604 (Figure 6) that directs the behavior of the I / O logic 604 in performing the interactive screening or monitoring conversation. Examples of dialogue actions include asking the patient a question, repeating a question, instructing the patient, politely ending the conversation, changing the display behavior of the patient device 112, and changing characteristics of the speech presented by the patient by the I / O logic 604, i.e., pace, volume, apparent gender of the voice, etc.

[0129] The interaction control logic generator 702 customizes interaction actions for the patient. The interaction control logic generator 702 receives data representing the patient's subjective preferences and the patient's clinical and social history from the screening or monitoring data store 210. In this exemplary embodiment, the subjective preferences are typically explicitly specified by the patient prior to an interactive screening or monitoring conversation and include such things as the particular voice to be presented to the patient through the I / O logic 604, the default volume and pace of speech generated by the I / O logic 604, and the display scheme to be used within the patient device 112.

[0130] The patient's clinical and social history, combined with the patient's identified interests, may indicate that questions related to particular topics should be asked of the patient. The dialogue control logic generator 702 uses the patient's preferences and medical history to set attributes of the questions to be asked of the patient.

[0131] The dialogue control logic generator 702 receives data representing the results of an analysis of the patient's responses in a current screening or monitoring conversation from the runtime model server logic 504. In particular, the dialogue control logic generator 702 receives data representing the results of the analysis of the responses, i.e., patient outcomes 1820 (FIG. 18), and result metadata from the descriptive model analysis 1812 that facilitates proper interpretation of the analysis results. The dialogue control logic generator 702 interprets the analysis results in conjunction with the result metadata to determine the patient's current status.

[0132] The History and State Machine 720 tracks the progress of the screening or monitoring conversation, i.e., which questions have been asked and which questions remain to be asked. The Question and Dialogue Action Bank 710 is a data store that stores all dialogue actions that can be taken by the Dialogue Control Logic Generator 702, including all questions that can be asked of the patient. Additionally, the History and State Machine 720 informs the Question and Dialogue Action Bank 710 as to which question is to be asked next in the screening or monitoring conversation.

[0133] The interaction control logic generator 702 receives data indicative of the current state of the conversation and which questions are queued to be asked from the history and state machine 720. The interaction control logic generator 702 processes the received data to determine the next action to be taken by the interactive screening or monitoring server logic 302 in facilitating the screening or monitoring conversation. Once the next action has been determined, the interaction control logic generator 702 retrieves data indicative of the action from the question and interaction action bank 710 and sends a request to the I / O logic 604 to perform the next action.

[0134] The overall implementation of a screening or monitoring conversation by the generic dialogue flow logic 602 is shown in logic flow diagram 800 (FIG. 8). The logic flow diagram of FIG. 8 describes the actions taken by the components of the dialogue engine in the block diagram of FIG. 28. Moreover, the logic flow diagram of FIG. 8 is an instantiation of the process described in FIG. 14. In step 802, the generic dialogue flow logic 602 selects a question or other dialogue action to initiate a conversation with the patient. The dialogue control logic generator 702 receives data from the history and state machine 720 indicating that the current screening or monitoring conversation is in its initial state. The dialogue control logic generator 702 receives data indicating (i) the subjective preferences of the patient and (ii) topics of relatively high relevance to the patient. Given that information, the dialogue control logic generator 702 selects an initial dialogue action with which the screening or monitoring conversation should be initiated. Examples of initial dialogue actions include (i) asking a general conversation-starting question such as "Can you hear me?" or "Are you ready to begin?", (ii) asking a question from a predefined script used with all patients, (iii) reminding the patient of a topic discussed in a previous screening or monitoring conversation with the patient and asking the patient a follow-up question related to that topic, or (iv) presenting the patient with several topics to choose from using user interface techniques at the patient device 112. In step 802, the dialogue control logic generator 702 causes the I / O logic 604 to perform the initial dialogue action.

[0135] The loop step 804 and the next stage step 816 define a loop in which the generic dialogue flow logic 602 conducts the screening or monitoring conversation according to steps 806-814 until the generic dialogue flow logic 602 determines that the screening or monitoring conversation is completed.

[0136] In step 806, the dialogue control logic generator 702 causes the I / O logic 604 to execute the selected dialogue action. In a first execution of step 806, a dialogue action is selected in step 802. In subsequent executions of step 806, a dialogue action is selected in step 814, as described below. In step 808, the generic dialogue flow logic 602 receives an audiovisual signal of the patient's response to the question. Although the process according to logic flow diagram 800 is shown in a manner that suggests synchronous processing, the generic dialogue flow logic 602 effectively executes step 808 continuously while steps 802-816 are being executed, processing the conversation asynchronously. The same is true for steps 810-814. In step 810, the I / O logic 604 transmits the audiovisual signal received in step 808 to the runtime model server logic 504, which processes the audiovisual signal in a manner described below. In step 812, the I / O logic 604 of the generic dialogue flow logic 602 receives the multiplex data from the runtime model server logic 504 and generates therefrom an intermediate score for the screening or monitoring conversation thus far.

[0137] As previously discussed, the outcome data includes analysis outcome data and outcome metadata. The I / O logic 604 (i) determines how complete the screening or monitoring conversation was in screening or monitoring the patient's target health condition(s), (ii) identifies any topics in the patient's responses that warrant follow-up questions, and (iii) identifies any explicit instructions from the patient to modify the screening or monitoring conversation. Examples of the latter include patient utterances such as "Can you speak louder?", "Can you repeat that?", or "What?", and "Speak slower." In step 814, the generic dialogue flow logic 602 selects the next question to ask the subject patient along with other dialogue actions to be performed by the I / O logic 604 in the next execution of step 806. In particular, the dialogue control logic generator 702 (i) receives dialogue state data from the history and state machine 720 regarding the next question to be asked, (ii) receives intermediate result data from the I / O logic 604 representing the patient's historical health status assessment, and (iii) receives the patient's preferences and relevant topics.

[0138] Processing transitions to loop step 804 via next step step 816. The generic dialogue flow logic 602 continues to loop through steps 804-816 until the dialogue control logic generator 702 determines that the screening or monitoring conversation is complete, at which point the generic dialogue flow logic 602 gracefully terminates the screening or monitoring conversation. A screening or monitoring conversation is complete when (i) all required questions have been asked and answered by the patient, and (ii) the confidence indicator of the score resulting from the screening or monitoring determined in step 812 is at least a predefined threshold. It should be noted that confidence in screening or monitoring is not symmetric.

[0139] The screening or monitoring conversation seeks to detect certain health conditions of the patient, e.g., depression and anxiety. If such conditions are detected immediately, they are detected. However, the absence of such conditions is not guaranteed by failure to detect them immediately. More generally, absence of evidence is not evidence of absence. Thus, the general dialogue flow logic 602 finds confidence in early detection, but not in early failure to detect. Thus, the health screening or monitoring server 102 (FIG. 4) estimates the patient's current health condition, e.g., mood, using a verbal conversation with the patient via the patient device 112. The interactive screening or monitoring server logic 502 transmits data representative of the resulting screening or monitoring of the patient to the patient's physician or other clinician by transmitting the data to the clinician device 114. Additionally, the interactive screening or monitoring server logic 502 records the resulting screening or monitoring in the screening or monitoring system data store 410. The first priority of the generic dialogue flow logic 602 is to elicit utterances from the patient that are highly informative regarding the health attributes for which the health screening or monitoring server 102 is screening the patient. For example, in this exemplary embodiment, the health screening or monitoring server 102 screens most patients for depression and anxiety. The analysis performed by the runtime model server logic 504 is most accurate when presented with patient utterances of a particular quality. In this context, utterance quality refers to the sincerity with which the patient is speaking. Generally speaking, high quality utterances are authentic and sincere, while low quality utterances come from patients who are not engaged in the conversation or are intentionally dishonest.

[0140] For example, if a patient does not care about the accuracy of the screening or monitoring and instead wants to answer all questions as quickly as possible and finish the screening or monitoring as soon as possible, it is unlikely to reveal much about the patient's true health. Similarly, if a patient tries to control the results of the screening or monitoring by giving false responses, not only will the responses be linguistically false, but the affective components of the utterances may be distorted or missing due to the patient's insincere engagement. There are several ways in which the generic dialogue flow logic 602 can increase the likelihood that the patient's responses are relatively highly informative. For example, the generic dialogue flow logic 602 may prompt the patient to engage with the interactive screening or monitoring server logic 502 as an audio diary whenever the patient is so inclined. Spontaneous utterances by the patient, whenever motivated, tend to be authentic and sincere, and therefore highly informative.

[0141] The generic dialogue flow logic 602 may also select topics that are relevant to the patient. These topics may include topics specific to the patient's clinical and social records, and topics specific to the patient's interests. Using topics that are of interest to the patient may negatively impact the patient's mood. For example, asking the patient about their favorite sports team may raise or lower the patient's mood with the latest news about the team. Thus, the generic dialogue flow logic 602 distinguishes health-related topics of interest to the patient from non-health-related topics of interest to the patient. For example, questions related to the patient's estranged relatives may be health-related, whereas questions related to the patient's favorite television series generally are not. Adapting any synthetic voice to match the patient's preferences may make the screening or monitoring conversation more appealing to the patient and thus elicit more informationally useful utterances. In embodiments in which the patient device 112 displays a video representation of a speaker, i.e., an avatar, to the patient, the patient's preferences include physical attributes of the avatar's appearance in addition to a preferred voice.

[0142] If the patient does not specify a preference for a synthetic voice or avatar, the generic dialogue flow logic 602 uses the synthetic voice and avatar selected for the initial screening or monitoring conversation and varies the synthetic voice and avatar in subsequent screening or monitoring conversations to compare the degree of informational usefulness of the patient's responses and determine which voice and avatar elicits the most informationally useful response. The voice and avatar selected for the initial screening or monitoring conversation may be selected according to which voice and avatar tend to elicit the most informationally useful utterances among the general population or among portions of the general population that share one or more phenotypes with the patient. The manner in which the informational usefulness of responses elicited by questions is determined is described below.

[0143] To make the screening or monitoring conversation more interactive and engaging, the generic dialogue flow logic 602 inserts a synthetic backchannel into the conversation. For example, the generic dialogue flow logic 602 may say "uh-huh" during short pauses in the patient's speech to suggest that the generic dialogue flow logic 602 is listening and interested in what the patient has to say. Similarly, the generic dialogue flow logic 602 causes a video avatar to exhibit non-verbal behaviors (sometimes called "body language") to indicate the patient's attention and interest.

[0144] The general dialogue flow logic 602 also selects quality questions.

[0145] The quality of a question is measured by the informativeness of the response it elicits. Additionally, the generic dialogue flow logic 602 avoids repeating the same question in subsequent screening or monitoring conversations, substituting equivalent questions where possible. The manner in which questions are determined to be equivalent to one another is described more fully below. As previously mentioned, the Question and Adaptive Action Bank 710 (FIG. 5) is a data store that stores all dialogue actions that can be taken by the dialogue control logic generator 702, including all questions that can be asked of a patient.

[0146] The Question and Adaptive Action Bank 710 is shown in more detail in FIG. 9. The Question and Adaptive Action Bank 710 is shown in more detail in FIG. 7 and includes several Question Records 902 and Dialogues 912. Each of the Question Records 902 includes data representing a single question that may be asked of a patient. The Dialogues 912 are a series of questions to be asked of a patient in a verbal conversation with the patient. Each of the Question Records 902 includes a Question Body 904, a Classification 906, a Quality 908, and an Equivalence 910. The Question Body 904 includes data specifying the substantive content of the question, i.e., the series of words to be spoken to the patient to ask the question. The Topic 906 includes data specifying a hierarchical topic category to which the question belongs. The categories may relate to (i) a particular medical diagnosis, such as depression, anxiety, etc.; (ii) a particular symptom, such as insomnia, lethargy, general apathy, etc.; and / or (iii) an aspect of the patient's treatment, such as medication, exercise, etc. Quality 908 includes data representative of the quality of the question. Question quality is an indication of the informativeness of the response elicited by the question. Equivalence 910 is data identifying one or more other questions in question records 902 that are equivalent to the question represented by this particular one of question records 902. In this exemplary embodiment, only questions of the same topic 906 may be considered equivalent. In another embodiment, any question may be considered equivalent regardless of classification. Dialogue 912 includes an ordered series of questions 914A-N, each of which identifies a respective one of question records 902 to ask in a verbal conversation with a patient. In this exemplary embodiment, the verbal conversation begins with twenty (20) preselected questions and may include further questions as necessary to generate a threshold confidence for terminating the conversation of logic flow diagram 600 (FIG. 6). The pre-selected questions include, in order, five (5) quality non-restrictive questions, eight (8) questions from the standard known PHQ-8 screening or monitoring tool for depression, and seven (7) questions from the standard known GAD-7 screening or monitoring tool for anxiety. In other examples, the questions may be generated algorithmically. In this exemplary embodiment, dialogue 912 specifies these twenty (20) questions.As discussed above with respect to step 814 (FIG. 10), the dialogue control logic generator 702 determines the next question to ask the patient in step 814. One embodiment of step 814 is shown as logic flow diagram 1014 (FIG. 8). In step 1002, the dialogue control logic generator 702 dequeues a question from the dialogue 912, which queues an ordered set of questions 914A-N. The history and state machine 720 keeps track of which of the questions 914A-N is next. If the screening or monitoring conversation is not completed according to the intermediate scores and all of the questions 914A-N have been processed in a previous execution of step 1002 in the same oral conversation, i.e., the queue of questions is empty, the dialogue control logic generator 702 selects a question from the questions in the question record 902 that has the highest quality 908 and is related to the topic selected for the patient.

[0147] If the dialogue control logic generator 702 selects multiple questions, the dialogue control logic generator 702 may randomly select questions to be dequeued, with each question weighted by its quality 908 and proximity to the suggested topic.

[0148] In step 1004 (FIG. 10), the dialogue control logic generator 702 collects all equivalent questions identified by equivalence 910 (FIG. 9) for the questions dequeued in step 1002. In step 1006, the dialogue control logic generator 702 selects a question from the set of equivalent questions collected in step 1004, including the question dequeued in step 1002 itself. The dialogue control logic generator 702 may randomly select one of the equivalent questions, or may use information about previous interactions with the patient, for example, to select one of the equivalent questions recently asked of the patient. The dialogue control logic generator 702 processes the selected question as the next question in the next iteration of the loop of steps 804-816 (FIG. 8). The use of equivalent questions is important. The quality of a question, i.e., the degree to which the responses it elicits are informationally useful in the runtime model server logic 504, will degrade over time for a given patient. In other words, if a given question is asked repeatedly to a given patient, each successive response by the patient will be less informative than it was in all previous questions asked. In a sense, questions become stale over time. To keep questions fresh, i.e., to solicit responses that are consistently informative over time, a given question is replaced in subsequent conversations with an equivalent but different question. However, the equivalence measure can be accurate with respect to comparing responses to questions that are consistently equivalent over time.

[0149] Thus, two important concepts of questions in the generic dialogue flow logic 602 (FIG. 7) are question quality and question equivalence. Question quality and question equivalence are managed by question management logic 916, which is shown in more detail in FIG. 11. Question management logic 916 includes question quality logic 1102 that measures question quality, and question equivalence logic 1104 that determines whether two (2) questions are equivalent in the context of the health screening or monitoring server 102. Question quality logic 1102 includes several metric records 1106 and metric aggregation logic 1112. To measure the quality of a question, i.e., how informative the responses elicited by a question are, question quality logic 1102 uses several metrics to be applied to the question, each of which results in a numerical quality score for the question, and each of the metrics is represented by one of the metric records 1106. Each of the metric records 1106 represents a single metric for measuring the quality of a question and includes metric metadata 1108 and quantification logic 1110. The metric metadata 1108 represents information about the metric of the metric record 1106. The quantification logic 1110 specifies the behavior of the question quality logic 1102 in assessing the quality of a question according to the metric of the metric record 1106.

[0150] The following are examples of metrics that may be applied by the question quality logic 1102 to measure the quality of various questions: (i) length of elicited response in terms of number of words, (ii) length of elicited response in terms of duration of response utterances, (iii) weighted word score, (iv) magnitude of acoustic energy in the elicited response, and (v) "vocal activation" in the response elicited by a question. Each will be described in turn.

[0151] In a metric record 1106 representing a metric of the length of the elicited response in terms of number of words, the quantification logic 1110 retrieves all responses to a given question from the screening or monitoring system data store 410 (FIG. 4) and determines the number of words in each response using associated results data from the screening or monitoring system data store 410. The quantification logic 1110 quantifies the quality of the question as a statistical indication of the number of words in the responses, e.g., as a statistical average thereof.

[0152] With respect to the length of an elicited response in terms of the duration of the response utterance, the duration of an elicited response may be measured in several ways. On the one hand, the duration of an elicited response is simply the elapsed duration, i.e., the entire duration of the response recorded in the screening or monitoring system data store 410. On the other hand, the duration of an elicited response is the elapsed duration minus any pauses in speech. In yet another, the duration of an elicited response is the elapsed duration minus any pauses in speech at the end of the response.

[0153] In a metric record 1106 (FIG. 11) representing a metric of the duration of an elicited response, the quantification logic 1110 retrieves all responses to a given question from the screening or monitoring system data store 410 (FIG. 4) and determines the duration of those responses. The quantification logic 1110 (FIG. 11) quantifies the quality of the question as a statistical indication of the duration of the responses, e.g., as a statistical average thereof.

[0154] With respect to the weighted word scores, the semantic model of the NLP model 1806 (FIG. 18) infers the patient's health condition from the positive and / or negative content of the patient's utterance. The semantic model associates individual words and phrases with the particular health condition that the semantic model is designed to detect. In a metric record 1106 (FIG. 11) representing a weighted word score metric, the quantification logic 1110 retrieves all responses to a given question from the collected patient data 410 (FIG. 5) and uses the semantic model to determine the correlation of each word of each response to one or more health conditions. The weighted word score of the individual responses is a statistical average of the correlations of the weighted word scores. The quantification logic 1110 quantifies the quality of the question as a statistical measure of the weighted word scores of the responses, e.g., as a statistical average thereof.

[0155] With respect to the magnitude of acoustic energy in the elicited responses, the runtime model server logic 504 (FIG. 18) estimates the patient's health status from the pitch and energy of the patient's speech, as described below. How informative an utterance is to the various models of the runtime model server logic 504 is directly related to how emotional the utterance is. In a metric record 1106 (FIG. 11) representing a metric of the magnitude of acoustic energy, the quantification logic 1110 retrieves all responses to a given question from the screening or monitoring system data store 410 (FIG. 4) and uses the response data from the runtime model server logic 504 to determine the magnitude of energy present in each response. The quantification logic 1110 quantifies the quality of the question as a statistical measure of the measured acoustic energy of the responses, e.g., as a statistical average thereof.

[0156] With respect to the "voice activation" of the responses elicited by a question, the quality of the question is a measure of how similar the response to the question is to the speech recognized by the runtime model 1802 (FIG. 18) as perfectly representative of the health condition that the runtime model 1802 is trained to recognize. In a metric record 1106 (FIG. 11) representing a voice activation metric, the quantification logic 1110 determines how similar the deep learning machine features in all responses to a given question are as a whole to the deep learning machine features in the health screening or monitoring server 102.

[0157] Deep learning machine features are known, but are briefly described herein to facilitate understanding and appreciation of the present invention. Deep learning is a sub-science of machine learning in that a deep learning machine is a machine learning machine, i.e., a learning machine that teaches itself how to distinguish one thing represented in data from other things represented in data. Below is a simple example to illustrate the difference.

[0158] Consider a regular (non-deep) learning machine configured to recognize the representation of a cat in image data. Such a learning machine is generally a computer process with multiple layers of logic. One layer is manually configured to recognize continuous portions of an image with transitions from one color to another (e.g., light to dark, red to green, etc.). This is commonly referred to as edge detection. Subsequent layers receive data representing the recognized edges and are manually configured to recognize edges that combine with each other to define a shape. The final layer receives data representing the shape and is manually configured to recognize symmetrical groupings of triangles (the cat's ears) and dark regions (the eyes and nose). Other layers may be used between the layers mentioned herein.

[0159] In machine learning, data received as input to any step in a computation, including intermediate results from other steps in the computation, are called features. The outcomes of the learning machine are called labels. In this illustrative example, the labels are "cat" and "no cat."

[0160] This manually constructed learning machine can work quite well, but it can have significant drawbacks. For example, recognizing symmetrical groupings of shapes may not recognize an image in which a cat is represented in profile. In a deep learning machine, the machine is trained to recognize a cat without manually specifying which group of shapes represents a cat. Although a deep learning machine may utilize manually constructed features to recognize edges, shapes, and groups of shapes, these are not required components of a deep learning system. Features in a deep learning system can be learned completely automatically by algorithms based only on labeled training data.

[0161] Training a deep learning machine to recognize cats in image data may involve, for example, presenting the deep learning machine with a large number of images, preferably millions of images, and associated knowledge regarding whether each image contains a cat, i.e., an associated indicator of "cat" or "no cat." For each image received in training, the final automatically constructed layer of the deep learning machine receives data representing a large number of groups of shapes and the associated indicators of "cat" or "no cat." Using statistical analysis and conventional techniques, the deep learning machine determines the statistical weight that each type of shape grouping, i.e., each feature, should be given in determining whether a previously unseen image contains a cat.

[0162] These trained, i.e., automatically generated, features of the deep learning machine likely include symmetric groupings of shapes that were manually configured into the learning machine as described above, but they also likely include shape groupings and combinations of shape groupings that the human programmer did not consider.

[0163] In measuring the quality of a question, the features of the constituent models of the runtime model server logic 504 (FIG. 18) specify precisely the types of responses indicative of health conditions that the constituent models of the runtime model server logic 504 are configured to recognize. Thus, these features represent a representative set of features when assessing the quality of a question. To measure the quality of a question using this metric, the quantification logic 1110 (FIG. 11) retrieves all responses to the question from the screening or monitoring system data store 410, as well as data representing diagnoses associated with those responses, and uses those responses and associated data to train the runtime model 1802 and the model repository 416.

[0164] In training the runtime model 1802 and the model repository 416, the deep learning machine generates a set of features that are specific to the question to be measured and the decision to be made by the trained model. The quantification logic 1110 measures the similarity between the question-specific feature set and the prototypical feature set in a manner described below with respect to the question equivalence logic 1104.

[0165] As previously discussed, the dialogue control logic generator 702 (FIG. 7) uses the various question qualities 908 (FIG. 9) in determining which question(s) to ask a particular patient. To provide a comprehensive indication of question quality to be stored in quality 908 (FIG. 9), the metric aggregation logic 1112 (FIG. 11) aggregates the various indications of quality according to the metric records 1106. The manner in which the aggregation logic 1112 aggregates the indications of quality of a given question is illustrated by logic flow diagram 1200 (FIG. 12).

[0166] Loop step 1202 and next step step 1210 define a loop in which metric aggregation logic 1112 processes each of the metric records 1106 according to steps 1204-1208. The particular one of the metric records 1106 processed in each iteration of the loop of steps 1202-1210 may be referred to as the "target metric record" and the metric represented by the target metric record may be referred to as the "target metric." In step 1204, metric aggregation logic 1112 evaluates the target metric using the target metric record quantification logic 1110 and all responses in the screening or monitoring system data store 410 (FIG. 4) to the target question. In test step 1206 (FIG. 12), metric aggregation logic 1110 determines whether the screening or monitoring system data store 410 contains a statistically significant sample of responses to the target questions by the target patients. If so, the metric aggregation logic 1110 evaluates the target metrics using the quantification logic 1110 and only the data corresponding to the target patient in the screening or monitoring system data store 410 in step 1208. Conversely, if the collected patient data 410 does not include a statistically significant sample of responses to the target questions by the target patient, the metric aggregation logic 1112 skips step 1208. Thus, the metric aggregation logic 1112 evaluates the quality of the questions in relation to the target patient to the extent the screening or monitoring system data store 410 includes sufficient data corresponding to the target patient.

[0167] After steps 1206-1208, processing continues via next step 1210 to loop step 1202 where metric aggregation logic 1110 processes the next metric according to the loop of steps 1202-1210. Once all metrics have been processed according to the loop of steps 1202-1210, processing continues to step 1212 where metric aggregation logic 1110 aggregates the evaluated metrics from all executions of steps 1204, 1206 into a single indicator of quality and stores data representative of that indicator of quality in quality 908. In this exemplary embodiment, metric metadata 1108 stores data that specifies how metric aggregation logic 1112 should include associated metrics in the aggregate indicator in step 1212. For example, metric metadata 1108 may specify the weight to attribute to associated metrics relative to other metrics.

[0168] After step 1212 (FIG. 12), processing for logic flow diagram 1200 is complete.

[0169] As previously described, the equivalence for a given question 910 identifies one or more other questions in the question records 902 that are equivalent to the given question. Whether two questions are equivalent is determined by the question equivalence logic 1104 (FIG. 11) by comparing the similarity between the two questions to a predefined threshold. Similarity in this case refers to how similar the models in the runtime model server 504 and the model repository 416 appear to be, not how similar the words or wording of the sentences are. The predefined threshold is determined empirically. The question equivalence logic 1104 measures the similarity between two questions in a manner illustrated by logic flow diagram 1300 (FIG. 13).

[0170] Loop step 1302 and next step step 1306 define a loop in which the question equivalence logic 1304 processes each of the metric records 1106 according to step 1304. The particular one of the metric records 1306 processed in each iteration of the loop of steps 1302-1306 may be referred to as the "target metric record" and the metric represented by the target metric record may be referred to as the "target metric." In step 1304, the question equivalence logic 1104 evaluates the target metric for each of the two questions. Once all metrics have been processed according to the loop of steps 1302-1106, processing by the question equivalence logic 1104 proceeds to step 1308.

[0171] In step 1308, the question equivalence logic 1104 combines the metrics evaluated for each question into a respective multi-dimensional vector for each question.

[0172] In step 1310, the query equivalence logic 1104 normalizes both vectors to have a length of 1.0. In step 1312, the query equivalence logic 1104 determines the angle between the two normalized vectors.

[0173] In step 1314, the cosine of the angle determined in step 1312 is determined by the query equivalence logic 1104 such that there is a measured similarity between the two queries.

[0174] Because vectors are normalized to a length of 1.0, the similarity between two questions ranges from -1.0 to 1.0, with 1.0 being perfect equivalence. In this exemplary embodiment, the pre-defined threshold is 0.98, such that two questions have a measured similarity of at least 0.98 and are considered equivalent, represented by equivalence 910 (FIG. 9) for both questions.

[0175] Furthermore, because the comparison between questions is a comparison of multidimensional vectors rather than a comparison of single values, two questions are equivalent not only if they are similar in general, but also if they are similar in most or all ways that are measured.

[0176] In another embodiment (FIG. 3), the assessment test administrator 2202 (FIG. 22) administers a depression assessment test to the subject patient by engaging in an interactive verbal conversation with the subject patient via the patient device 312. The manner in which the assessment test administrator 2202 does so is illustrated in logic flow diagram 1400 (FIG. 14). The test administrator 2202 may be a computer program configured to ask questions of the patient. The questions may be algorithmically generated questions. The questions may be generated, for example, by natural language processing (NLP) algorithms. Examples of NLP algorithms are semantic parsing, sentiment analysis, vector space semantics, and relationship extraction. In some embodiments, the methods described herein may be capable of generating an assessment without the need for the presence or intervention of a human clinician. In other embodiments, the methods described herein may be used to augment or enhance the assessment provided by the clinician or to assist the clinician in providing the assessment. The assessment may include queries that include themes adapted or modified from screening or monitoring methods, such as the PHQ-9 assessment and the GAD-7 assessment. The evaluation herein may not only use questions from such surveys verbatim, but may also adaptively modify queries based at least in part on responses from subject patients.

[0177] In step 1402, the evaluation test administrator 2202 optimizes the test environment. Step 1402 is illustrated in further detail in logic flow diagram 1402 (FIG. 15).

[0178] In step 1502, the assessment test administrator 2202 begins a verbal conversation with the target patient. In this exemplary embodiment, the assessment test administrator 2202 begins the conversation by asking the patient the first question of the assessment test. The first question is selected in a manner that is more fully described below. The exact question asked is not particularly important. What is important is that the patient responds with enough speech to enable the assessment test administrator 2202 to evaluate the quality of the video and audio signals received from the patient device 312.

[0179] The assessment test administrator 2202 receives and processes audiovisual data from the patient device 312 throughout the conversation. Loop step 1504 and next step step 1510 define a loop in which the assessment test administrator 2202 processes the audiovisual signal according to steps 1506-1508 until the assessment test administrator 2202 determines that the audiovisual signal is of high quality, or at least has a quality suitable for making an accurate assessment.

[0180] In step 1506, the assessment test administrator 2202 evaluates the quality of the audiovisual signals received from the patient device 312. In particular, the assessment test administrator 2202 measures the volume of speech, the clarity of speech, and the extent to which the patient's face and, if available, body can be seen.

[0181] In step 1508, the assessment test administrator 2202 reports the assessment to the patient. In particular, the assessment test administrator 2202 generates an audiovisual signal representing a message to be played to the patient via the patient device 312. If the audiovisual signal received from the patient device 312 is determined by the assessment test administrator 2202 to be of insufficient quality, the message asks the patient to adjust their environment to improve the signal quality. For example, if the audio portion of the signal is insufficient, the message may be "I'm having trouble hearing you. Can you move the microphone closer to you or find a quieter place?", and if the patient's face and body, if available, cannot be clearly seen, the message may be "I can't see your face (and body). Can you reposition your phone so I can see you?" After step 1508, processing by the assessment test administrator 2202 proceeds to loop step 1504 via next step step 1510, where the assessment test administrator 2202 continues processing according to the loop of steps 1504-1510 until it is determined that the received audiovisual is adequate or as good as it can be obtained for the current assessment test. A subsequent execution of step 1508 is preferably in response to any speech by the patient. For example, the patient may respond to the message to improve the environment with the question "Is this better?". The next message sent in the report of step 1508 should include the answer to the patient's question. As described herein, the composite model 2204 includes a language model component, and therefore the assessment test administrator 2202 necessarily performs speech recognition.

[0182] Once it is determined that the received audiovisual quality is adequate or as good as it can be for the current assessment test, processing by assessment test administrator 2202 according to the loop of steps 1504-1510 is complete. Furthermore, processing according to logic flow diagram 1402, and thus step 1402 (FIG. 14), is complete.

[0183] Loop step 1404 and next step step 1416 define a loop in which assessment test administrator 2202 performs the oral dialogue of the assessment test according to steps 1406-1414 until assessment test administrator 2202 determines that the assessment test is completed.

[0184] In step 1406, the assessment test administrator 2202 asks the patient questions to facilitate verbal conversation. In this exemplary embodiment, the assessment test administrator 2202 asks the patient using queued questions, which may also be referred to herein as the conversation queue. In a first run of step 1406, the queue may be pre-populated with questions to be covered during the assessment test. Generally, these questions cover the same general subject matter covered by currently used written assessment tests, such as the PHQ-9 and GAD-7. However, while the questions in these tests are purposefully designed to elicit very short, direct answers, the assessment test administrator 2202 may require more audio and video than is provided by a one-word answer. Therefore, it is preferable that the questions initially queued are more non-restrictive.

[0185] In this exemplary embodiment, the initial questions relate to the topics of general mood, sleep, and appetite. An example of an initial question related to sleep is question 1702 (FIG. 17), i.e., "How have you been sleeping lately?" This question is intended to elicit a sentence or two from the patient, thereby providing more audio and video of the patient than would normally be elicited by a very direct question.

[0186] In step 1408 (FIG. 14), the assessment test administrator 2202 receives audiovisual signals of the patient's responses to the questions. Although the process according to logic flow diagram 1400 is shown in a manner that suggests synchronous processing, the assessment test administrator 2202 effectively continues to perform step 1408 while steps 1402-1416 are being performed, processing the conversation asynchronously. The same is true with respect to steps 1410-1414.

[0187] In step 1410, the assessment test administrator 2202 processes the audiovisual signals received in step 1408 using the composite model 2204. In step 1412, the assessment test administrator 2202 generates an intermediate score for the assessment test according to the audiovisual signals received so far.

[0188] In step 1414, the assessment test administrator 2202 selects the next question to be asked of the subject patient in the next execution of step 1406, and processing transitions to next stage step 1416 and loop step 1404. Step 1414 is shown in more detail as logic flow diagram 1414 (FIG. 16). Additionally, FIG. 16 may be interpreted as continuing from step 814 of FIG.

[0189] In step 1602, the assessment test administrator 2202 identifies significant elements in the patient's speech. In particular, the assessment test administrator 2202 uses the linguistic portion of the composite model 2204 to identify distinct assertions in the portion of the audiovisual signal received after the last question was asked in step 1406 (FIG. 14). That portion of the audiovisual signal may be referred to herein as the "patient's response" in the context of a particular iteration of the loop of steps 1604-1610.

[0190] An example of a conversation conducted by the assessment test administrator 2202 of the real-time system 302 and the patient device 312 is shown in (FIG. 17). It should be understood that the conversation 1700 is merely exemplary. The particular questions to be asked, what parts of the patient's response are significant, and the depth to which any topics are continued are determined by and configured in the type information to be collected by the assessment test administrator 2202. In step 1702, the assessment test administrator 2202 asks the question, "How are you sleeping lately?" The patient's response is, "Okay...I've been having a hard time sleeping lately. I have medicine for that. They seem to be helping." In step 1602, the assessment test administrator 2202 identifies three (3) significant elements in the patient's response: (i) "sleep disturbance," which suggests the patient has some form of insomnia or at least is not sleeping enough, (ii) "I have medicine," which suggests the user is taking medicine, and (iii) "they seem to be helping," which suggests the medicine taken by the user is effective. In the illustrative example of the conversation 1700, each of these significant elements is processed by the assessment test administrator 2202 in a loop of steps 1604-1610.

[0191] Loop step 1604 and next step step 1610 define a loop in which the assessment test administrator 2202 processes each significant element of the patient's answers identified in step 1602 according to steps 1606-1608. In the context of a given iteration of the loop of steps 1604-1610, the particular significant element being processed may be referred to as a "target element." In step 1606, the assessment test administrator 2202 processes the target element, thereby recording details contained in the element and identifying follow-up questions. For example, in conversation 1700 (FIG. 17), the assessment test administrator 2202 identifies three (3) topics for follow-up questions regarding the insomnia element: (i) type of insomnia (early, intermediate, or late), (ii) frequency of insomnia experienced by the patient, and (iii) what medication, if any, is being taken by any patient for insomnia.

[0192] In step 1608 , the assessment test administrator 2202 queues the follow-up questions identified in step 1606 .

[0193] After step 1608, processing by the assessment test administrator 2202 proceeds to loop step 1604 via next step step 1610 until the assessment test administrator 2202 has processed all significant elements of the patient's responses in accordance with the loop of steps 1604-1610, at which point processing proceeds from loop step 1604 to step 1612.

[0194] In the exemplary context of conversation 1700 (FIG. 17), the state of the conversation queue is as follows: FIG. 17 illustrates a particular instantiation of a conversation going on between the system and the patient. The queries and replies disclosed herein are exemplary and should not be construed as necessarily following the order disclosed in FIG. 17. When processing the insomnia response element, the assessment test administrator 2202 identifies and queues follow-up topics regarding the type of insomnia and any medications taken for insomnia.

[0195] In processing the medication response element in the patient's response, the assessment test administrator 2202 finds that the patient is taking medication, and in step 1606, the assessment test administrator 2202 records that fact and identifies a queued follow-up question regarding medication for insomnia and processes the medication element in response to the queued question.

[0196] In step 1608 for the medication element, the assessment test administrator 2202 queues follow-up questions regarding the specific medication and dosage used by the patient and its effectiveness, as shown in step 1708.

[0197] In this exemplary embodiment, the questions in the conversation queue are hierarchical. In the hierarchy, each follow-up question is a child of the question it follows up on. The latter question is the parent of the follow-up question. When removing a question from the conversation queue, the assessment test administrator 2202 performs a first-order depth-first walk of the conversation queue hierarchy. In other words, all child questions of a given question are processed before processing the next sibling question. In conversational terms, all follow-up questions of a given question are processed before recursively processing the next question at the same level. In the context of the conversation 1700, the assessment test administrator 2202 processes all follow-up questions for insomnia type before processing the frequency and dosage questions and any of their follow-up questions. This is the way conversations naturally occur - staying on the last topic discussed until completion before returning to a previously discussed topic.

[0198] Additionally, the order in which sibling questions are processed by the assessment test administrator 2202 may be influenced by the patient's response. In this illustrative example, the follow-up question regarding the frequency of insomnia precedes the follow-up question regarding medication. However, when processing the medication element in step 1606, the assessment test administrator 2202 changes the order of the follow-up questions so that the follow-up question regarding medication is processed before processing the follow-up question regarding the frequency of insomnia. Because medication is mentioned by the patient, it is explained before adding a new side topic to the conversation. This is another way in which the assessment test administrator 2202 can respond to the patient.

[0199] When processing the efficacy response element (i.e., "these medicines seem to be helping"), the assessment test administrator 2202 records that the drug is moderately effective. Upon seeing that the conversation queue contains a question about efficacy, the assessment test administrator 2202 applies this portion of the patient's response as a response to the queued follow-up question in step 1710.

[0200] In step 1612, the assessment test administrator 2202 removes the next question from the conversation queue, the process according to logic flow diagram 1414, and therefore step 1414, is completed, and the conversation continues. Before returning to the discussion of (FIG. 14), it is useful to consider the further execution of step 1414, and therefore logic flow diagram 1414, in the context of the exemplary conversation 1700. The question removed from the queue as the next question in this exemplary embodiment asks about the patient's insomnia, thereby attempting to identify the type of insomnia. It can be seen that the conventional thinking reflected in the PHQ-9 and GAD-7 is that the particular type of sleep disturbance experienced by the subject is not as strong an indicator of depression as the mere fact that it is difficult to sleep. However, delving deeper into the topic of conversation has several beneficial consequences. Most importantly, the user is encouraged to provide more speech to more accurately assess the patient's condition. Furthermore, by asking questions about what the patient has just said, the assessment test administrator 2202 suggests that he or she takes a personal interest in the patient and, by gaining goodwill from the patient, makes the patient more likely to be honest in both speech and behavior.

[0201] In the illustrative example of conversation 1700, the next question is associated with the type of insomnia. The question is intentionally as non-restrictive as possible while still targeting specific information, i.e., "Do you ever wake up in the middle of the night?" See question 1712. This question may elicit a "yes" or "no" answer, but may also elicit a longer response such as response 1714, i.e., "No. I just have trouble falling asleep." After step 1612, processing according to logic flow diagram 1414, and thus step 1414 (FIG. 14), is completed. In successive iterations of the loop of steps 1404-1416, assessment test administrator 2202 continues the illustrative example of conversation 1700. In the next execution of step 1406, assessment test administrator 2202 asks question 1712 (FIG. 17). In the next continued execution of step 1408, assessment test administrator 2202 receives response 1714. The assessment test administrator 2202 processes the response 1714 in the next execution of step 1414 .

[0202] In this exemplary implementation of step 1602 (FIG. 16), the assessment test administrator 2202 identifies a single significant factor: the patient has trouble falling asleep and not waking up during the night. In step 1606, the assessment test administrator 2202 records the type of insomnia (see step 1716), which in this exemplary embodiment has no associated follow-up questions.

[0203] In this example implementation of step 1608, the assessment test administrator 2202 removes the next question from the conversation queue. Because the follow-up questions about the type of insomnia and whether the patient is treating the insomnia with medication have not yet been answered, the next question is a first child question related to medication, i.e., the specific medication taken by the patient.

[0204] In the next iteration of step 1406 (FIG. 14), the assessment test administrator 2202 formulates a question, namely, what particular medication the patient is taking for insomnia. In some embodiments, the assessment test administrator 2202 asks the question in the most direct way, for example, "You said you are taking medication for insomnia. What medication are you taking?" This has the advantage of being less restrictive and eliciting more speech than a simple yes / no question.

[0205] In other embodiments, the assessment test administrator 2202 accesses clinical data associated with the patient to help identify the specific medication used by the patient. The clinical data may be received from the modeling system 302 (FIG. 22) or from the clinical data server 306 (FIG. 3) using clinical data 2220. Thus, the assessment test administrator 2202 may ask a more direct question using the most common and generic name of the assumed medication. For example, if the patient's data indicates that the patient is prescribed Zolpidem (the generic name of a drug sold under the brand name Ambien), then the question 1720 (FIG. 17) may be "You mentioned that you are taking a medication for insomnia. Is it Ambien or Zolpidem?" This very straightforward question risks eliciting only a simple yes / no response (e.g., response 1722). However, the question also demonstrates knowledge of and interest in the patient, thereby gaining further goodwill and increasing the likelihood of an honest response by the patient and willingness to continue the assessment test for longer.

[0206] In this exemplary embodiment, the assessment test administrator 2202 determines whether to ask very straightforward questions, rather than more non-restrictive questions, based on whether the necessary clinical data is available for the patient and how much further speech is needed to obtain an adequate degree of accuracy in assessing the patient's condition.

[0207] The illustrative example of the conversation 1700 continues with the assessment test administrator 2202 recording the substance of the response 1722 at step 1724 .

[0208] The assessment test administrator 2202 in this exemplary embodiment also determines responsiveness to the patient in that, for example, when determining when the answers received in step 1408 are complete and selection of the next question in step 1414 can begin, the assessment test administrator 2202 determines whether the patient has completed his or her response to the most recently asked question.

[0209] To further promote goodwill among the patient, the assessment test administrator 2202 attempts to interrupt the patient as little as possible. Response 1704, i.e., "Okay.... I've been having trouble sleeping lately. I have medicine for that. They seem to be helping." The ellipsis after "Okay" indicates a pause in the patient's response. To this end, the assessment test administrator 2202 waits long enough to allow the patient to pause momentarily without interruption, but does not wait so long that it makes the patient believe that the assessment test administrator 2202 has become unresponsive, for example, due to a failure of the assessment test administrator 2202 or the communication link therewith. Additionally, speech pauses are used in the assessment as described more fully below, and the assessment test administrator 2202 should avoid disrupting the patient's speech fluency.

[0210] In this exemplary embodiment, the assessment test administrator 2202 uses two pause durations: a short pause duration and a long pause duration. After pausing for the short pause duration, the assessment test administrator 2202 indicates that the assessment test administrator 2202 continues to listen by playing a very short sound that acknowledges understanding and continuation of listening, such as "yes" or "hmm, yes." After playing the message, the assessment test administrator 2202 waits during any continuing pause for the long duration. If the pause continues that long, the assessment test administrator 2202 determines that the patient has completed their response.

[0211] The particular respective lengths of the short and long durations may be determined empirically. Furthermore, the optimal lengths may vary from patient to patient. Thus, the assessment test administrator 2202 continues to adjust these durations for the patient whenever interacting with the patient. The assessment test administrator 2202 recognizes a duration that is too short when it notices crosstalk, i.e., when the assessment test administrator 2202 is simultaneously playing any audio speeds received from the patient. The assessment test administrator 2202 recognizes a duration that is too long (i) when the patient explicitly indicates so (e.g., by saying "hello?" or "are you there yet?") and / or (ii) when the patient's reaction suggests increasing frustration or arousal at the patient's utterance earlier in the same conversation.

[0212] The conversation is gracefully terminated by the assessment test administrator 2202 when the assessment test is completed. The assessment test is completed when (i) the first question in the conversation queue and all of their descendants have been answered by the patient, or (ii) the indicator of confidence in the score resulting from the assessment determined in step 1412 is at least a predefined threshold. It should be noted that confidence in the assessment is not symmetric. The assessment test looks for depression or other behavioral health conditions in the patient. If it is found quickly, it is found. However, not finding it quickly does not guarantee its absence. Thus, the assessment test administrator 2202 finds confidence in early detection, but not in early detection failure.

[0213] In this manner, the real-time system 302 (FIG. 22) assesses the patient's current mental state using interactive verbal conversation with the patient via the patient device 312. The assessment test administrator 2202 transmits data representing the patient's resulting assessment to the patient's physician or other clinician by transmitting the data to the clinician device 314. Additionally, the assessment test administrator 2202 records the resulting assessment in clinical data 2220.

[0214] Although the assessment test administrator 2202 is described as conducting an interactive verbal conversation with the patient to assess the patient's mental state, in other embodiments, the assessment test administrator 2202 passively listens to the patient talking to a clinician and evaluates the patient's speech in the manner described herein. The clinician may be a mental health professional, a general practitioner, or a specialist such as a dentist, cardiac surgeon, or ophthalmologist. In one embodiment, the assessment test administrator 2202 passively listens to the conversation between the patient and the clinician via the patient device 312 when it determines, for example, by a "start" control on the clinician's iPad, that the patient is conversing with the clinician. The assessment test administrator 2202 stops passively listening and evaluating the speech in the manner described above when it determines, for example, by a "stop" control on the clinician's iPad, that the conversation between the patient and the clinician is completed. Further, because the patient device 312 listens passively and does not prompt the patient, the assessment test administrator 2202 makes no attempt to optimize the audiovisual signals received via the patient device 312, nor does it assume that faces in any received video signals are those of the patient.

[0215] In some embodiments, at the start of a conversation between the patient and clinician, the clinician asks the patient to begin listening with the assessment test administrator 2202, and the patient does so by issuing a command via the patient device 312 instructing the assessment test administrator 2202 to begin listening. Similarly, at the end of the conversation, the clinician asks the patient to end listening with the assessment test administrator 2202, and the patient does so by issuing a command via the patient device 312 instructing the assessment test administrator 2202 to stop listening.

[0216] In another embodiment, the assessment test administrator 2202 listens to the conversation between the patient and the clinician via the clinician device 314. The clinician may manually start and stop listening by the assessment test administrator 2202 via the clinician device 314 using conventional user interface techniques.

[0217] During conversations passively listened to by the assessment test administrator 2202, the assessment test administrator 2202 evaluates the patient's speech, not the clinician's speech. The assessment test administrator 2202 may distinguish between the voices in any of several ways, for example, by a "mute" control on the clinician's iPad. In embodiments where the assessment test administrator 2202 listens via the patient device 312, the assessment test administrator 2202 uses an acoustic model (e.g., acoustic model 2218) to distinguish between the two voices. The auxiliary test administrator 2202 identifies the louder voice as the patient's voice, assuming that the patient device 312 is closer to the patient than the clinician. This may also be the case in embodiments where the clinician device 312 is configured to listen to the patient more loudly. For example, the clinician device 314 may be configured to listen via a highly directional microphone that the clinician directs towards the patient such that the captured audio signal represents the patient's voice much more loudly than other ambient sounds, such as the clinician's voice. The assessment test administrator 2202 may further distinguish the patient's voice from the clinician's voice using a language model 2214, and in particular a semantic pattern model such as the semantic pattern module 4004, to identify which of the two distinguished voices asks the question more frequently. The assessment test administrator 2202 may further distinguish the patient's voice from the clinician's voice using an acoustic model 2016, which may identify and segment the clinician's voice from an acoustic analysis of the clinician's voice performed prior to the clinical encounter.

[0218] Throughout the conversation between the patient and the clinician, the assessment test administrator 2202 evaluates the patient's mental state from the patient's speech in the manner described herein and terminates the evaluation upon detecting the conclusion of the conversation.

[0219] The runtime model server logic 704, shown in more detail in FIG. 18, processes audiovisual signals representing a patient's responses in an interactive screening or monitoring conversation and infers the patient's current health state from the audiovisual signals while the conversation is ongoing.

[0220] The automatic speech recognition (ASR) logic 1804 is logic that processes the speech represented in the audiovisual data from the I / O logic 604 (FIG. 6) to identify words spoken in the audiovisual signal. The results of the ASR logic 1804 (FIG. 18) are sent to the runtime model 1802.

[0221] The runtime model 1802 also receives audiovisual signals directly from the I / O logic 604. In a manner described more fully below, the runtime model 1802 combines the language model, the acoustic model, and the visual model to generate results 1820 from the received audiovisual signals. The interactive screening or monitoring server logic 702 then uses the results 1820 in real time, as described above, to estimate the patient's current state and respond accordingly with verbal dialogue to the patient, as described above.

[0222] In addition to identifying words in the audiovisual signal, the ASR logic 1804 also identifies where each word appears in the audiovisual signal and a confidence in the accuracy of each identified word in this exemplary embodiment. The ASR logic 1804 may also identify non-verbal content of the audiovisual signal, such as laughter and fillers, along with the location and confidence information. The ASR logic 1804 makes such information available to the runtime model 1802.

[0223] The runtime models 1802 include a descriptive model and analysis 1812, a natural language processing (NLP) model 1806, an acoustic model 1808, and a visual model 1810.

[0224] The NLP models 1806 include several text-based machine learning models for (i) predicting depression, anxiety, and possibly other health conditions directly from words spoken by the patient, and (ii) modeling factors that correlate with such health conditions. Examples of machine learning models that directly model health conditions include sentiment analysis, semantic analysis, language modeling, word / document embedding and clustering, topic modeling, discourse analysis, syntactic analysis, and dialogue analysis. The models need not be constrained to one type of information. The models may include information from both emotion-based and topic-based features, for example. NLP information includes score outputs of specific modules, for example scores from an emotion detector that is trained on emotions and not on mental health conditions. NLP information includes information obtained via transfer learning-based systems.

[0225] The NLP model 1806 stores text metadata and modeling dynamics and shares that data with the acoustic model 1808, the visual model 1810, and the descriptive model and analysis 1812. The text data may be received directly from the ASR logic 1804 as described above, or may be received as text data from the NLP model 1806. The text metadata may include, for example, data identifying the part of speech (syntactic analysis), sentiment analysis, semantic analysis, topic analysis, etc. for each word or phrase. The modeling dynamics includes data representing components of the constituent models of the NLP model 1806. Such components include the machine learning features of the NLP model 1806 and other components such as long short-term memory (LSTM) units, gated recurrent units (GRUs), hidden Markov models (HMMs), sequence-to-sequence (seq2seq) transformation information, etc. The NLP metadata enables the acoustic model 1808, the visual model 1810, and the descriptive model and analysis 1812 to associate syntactic, emotional, semantic, and topical information with corresponding portions of the audiovisual signal, thus enabling the acoustic model 1808, the visual model 1810, and the descriptive model and analysis 1812 to more accurately monitor the audiovisual signal.

[0226] The runtime model 1802 includes an acoustic model 1808 that analyzes the audio portion of the audiovisual signal to find patterns associated with various health conditions, e.g., depression. The association between acoustic patterns in speech and health can potentially be applied to different languages ​​without retraining. These languages ​​may also be retrained based on data from the languages. A of a particular language is spoken. Thus, the acoustic model 1808 analyzes the audiovisual signal in a language-agnostic manner. In this exemplary embodiment, the acoustic model 1808 uses machine learning techniques such as convolutional neural networks (CNNs), long short-term memory (LSTM) units, and hidden Markov models (HMMs) to learn high-level representations and to model the temporal dynamics of the audiovisual signal.

[0227] The acoustic model 1808 stores data representing attributes of the audiovisual signal and machine learning features of the acoustic model 1808 as acoustic model metadata and shares the data with the NLP model 1806, the vision model 1810, and the descriptive model analysis 1812. The acoustic model metadata may include, for example, data representing a spectrogram of the audiovisual signal of the patient's response. Additionally, the acoustic model metadata may include both basic and higher level feature representations of the machine learning features. More basic features may include Mel-Frequency Cepstral Coefficients (MFCCs) and, for example, various logarithmic filter banks of the acoustic model 1808. High level feature representations may include, for example, convolutional neural networks (CNNs), autoencoders, variational autoencoders, deep neural networks, and support vector machines of the acoustic model 1808. The acoustic model metadata may enable the NLP model 1806 to improve sentiment analysis of words and phrases using, for example, acoustic analysis of the audiovisual signal. The acoustic model metadata allows the visual model 1810 and the descriptive model analysis 1812 to more accurately model the audiovisual signal, for example, using an acoustic analysis of the audiovisual signal.

[0228] The runtime model server logic 504 (FIG. 18) includes a visual model 1810 that infers various patient health conditions from facial, gaze, and posture behavior. The visual model 1810 may include facial cue modeling, eye / gaze modeling, posture tracking and modeling, etc., these are merely examples.

[0229] The visual model 1810 stores data representing attributes of the audiovisual signals and machine learning features of the visual model 1810 as visual model metadata and shares the data with the NLP model 1806, the acoustic model 1808, and the descriptive model and analysis 1812. For example, the visual model metadata can include data representing face position, posture tracking information, and gaze tracking information of the audiovisual signals of the patient's response. Additionally, the visual model metadata can include both basic and high level feature representations of the machine learning features. The more basic features can include the image processing features of the visual model 1810. The high level feature representations can include, for example, CNNs, autoencoders, variational autoencoders, deep neural networks, and support vector machines of the visual model 1810. The visual model metadata can enable the descriptive model and analysis 1812 to improve sentiment analysis of words and phrases using, for example, video analysis of the audiovisual signals. The descriptive model analysis 1812 can even use the visual model metadata in combination with the acoustic model metadata to estimate the patient's veracity in speaking words and phrases for more accurate sentiment analysis. The visual model metadata enables the acoustic model 1808 to use, for example, video analysis of the audiovisual signal to better interpret acoustic signals associated with various gazes, postures, and gestures depicted in the video portion of the audiovisual signal.

[0230] Descriptive features or descriptive analysis are interpretable descriptions that can be computed based on features in speech, language, video, and metadata that convey information about a speaker's speech patterns in a way that stakeholders can understand. For example, descriptive features can include a speaker who seems nervous or anxious, who has a high-pitched or deep voice, or who speaks quickly or slowly. Humans can interpret voice "features" such as pitch, rate, and semantics of speech to mentally determine emotions. A descriptive analysis module differs from a machine learning module by applying interpretable labels to speech based on those features. Although machine learning models also make predictions by analyzing features, the way machine learning algorithms process features and determine the display of those features differs from the way humans interpret them. Thus, the labels that machine learning algorithms can "apply" to data in the context of analyzing features need not be human-interpretable labels.

[0231] Descriptive model analysis 1812 (FIG. 18) may generate analyses and indicators for many health conditions, not just depression. Examples of such indicators include mood, anxiety, patient engagement, patient energy, emotion, speech rate, and topic of conversation. Descriptive model analysis 1812 then applies these indicators to each word of the patient's response to determine how significant each word is in the patient's response. While the significance of any given word in a verbal response may be inferred from the part of speech, e.g., relatively insignificant articles and filler words, descriptive model analysis 1812 infers the significance of a word from additional qualities of the word, such as the emotion in the way the word is spoken as indicated by acoustic model 1808.

[0232] The descriptive model analysis 1812 may also analyze trends over time and use such trends, at least in part, to normalize the analysis of the patient's responses. For example, a given patient may generally speak with less energy than other patients. A normalization analysis for this patient may set a lower level of energy as "normal" than would be used for the general population. Additionally, a given patient may use certain words more frequently than the general population, and the use of such words by this patient may not be as pronounced as such use by a different patient. The descriptive model analysis 1812 may analyze trends in real time, i.e., while a screening or monitoring conversation is ongoing, as well as in non-real time situations.

[0233] Descriptive model analysis 1812 stores data representing the speech analysis and trend analysis, as well as metadata of its constituent models, as descriptive model metadata, and shares that data with NLP model 1806, acoustic model 1808, and visual model 1810. The descriptive model metadata enables NLP model 1806, acoustic model 1808, and visual model 1810 to more accurately monitor the audiovisual signals.

[0234] Via the runtime model 1802, the runtime model server logic 504 estimates the patient's health state using a combination of what the patient says, how the patient says it, and contemporaneous facial expressions, eye looks, and posture, and stores the result data representative of such estimate as results 1820. Such provides a particularly accurate and effective tool for estimating the patient's health state.

[0235] The runtime model server logic 504 sends the results 1820 to the I / O logic 604 (FIG. 6) so that the interactive screening or monitoring server logic 502 can respond to the patient's responses, such that the screening or monitoring dialogue is interactive in the manner described above. The runtime model server logic 504 (FIG. 18) also sends the results 1820 to the screening or monitoring system data store 410 for inclusion in the subject's history.

[0236] The model training logic 506, shown in more detail in FIG. 19, trains the models used by the runtime model server logic 504 (FIG. 18).

[0237] The model training logic 506 ( FIG. 19 ) includes a runtime model 1802 and ASR logic 1804, and trains the runtime model 1802. The model training logic 506 sends the trained model to the model repository 416 and makes the trained runtime model 1802 available to the runtime model server logic 504.

[0238] FIG. 20A provides a more detailed illustrative example of the back-end screening or monitoring system of the embodiment of FIG. 2. In this example of a block diagram, the web server 240 is expanded to show that it includes a group of functional modules. The main components of the web server 240 include an input / output (IO) module 2041 for accessing the system via a network infrastructure 250. This IO 2041 enables the collection of response data (at least in the form of speech and video data) and indicators from the clients 260a-n, the presentation of prompt information (such as questions or topics), and feedback to the clients 260a-n. The prompt material is provided by a dialogue engine 2043 that responds to the system's needs, user commands, and preferences to build a dialogue that keeps the clients 260a-n engaged and generates meaningful response data. The dialogue engine is described in more detail below.

[0239] Patient truthfulness in answering questions (or other forms of dialogue) posed by a screening or monitoring test is important in assessing the patient's mental state, as is having a system that is familiar and sought after and used by prospective patients. The health screening or monitoring system 200 promotes patient truthfulness in several ways. First, a verbal conversation requires less time for the patient to formulate a response to a question or talk about a topic than a written response would require. This truncated time generally results in more truthful and "live" responses. Second, the conversation feels more spontaneous and personal to the patient than an apparently generic questionnaire, especially when user preferences are factored into the dialogue, as described below. Thus, the verbal conversation does not induce or amplify the patient's resentment about having to answer a questionnaire before seeing a doctor or other clinician. Third, the verbal conversation is adapted on the fly to respond to the patient, thereby reducing the patient's discomfort with the screening or monitoring test and, in some cases, shortening the screening or monitoring test. Fourth, the screening or monitoring tests performed by the health screening or monitoring system 200 rely on more than just the verbal components of the dialogue. Non-verbal aspects of the dialogue are exploited synergistically with the verbal content to assess a patient's depression. In effect, what is said is not as reliable and accurate in assessing depression as what is said.

[0240] The final component of the web server 240 is a results and presentation module 2045 that collates results from the model server(s) 230 and provides them to the clients 260a-n via the IO 2041, as well as providing feedback information to the dialogue engine 2043 for dynamically adapting the course of the dialogue to achieve the system's goals. Additionally, the results and presentation module 2045 further provides filtered results to the stakeholders 270a-n via the stakeholder communication module 2003. The communication module 2003 encompasses a process engine, a routing engine, and a rules engine. The rules engine embodies the conditional logic that determines what, when, and to whom to send a communication, the process engine embodies the clinical and computational protocol logic to send messages through a communication chain that may be based on the sequential completion of tasks, and the routing engine provides the ability to send any message to the user's optimal platform (e.g., mobile phone, computer, landline, tablet, etc.).

[0241] Filtering and / or modifying of results by the Results and Presentation Module 2045 is performed as necessary to present the most useful information to relevant stakeholders 270a-n while maintaining HIPAA (Health Insurance Portability and Accountability Act of 1996) and other privacy and security policies and regulations, such as GDPR and SOC2 compliance, as necessary. For example, a clinician may want to receive not only the screening or monitoring classification (e.g., depressed or neurotypical), but also further descriptive features such as suicidal ideation, anxiety on other topics, etc. In contrast, an insurance company may not need or want many of these further features and may only be interested in the diagnostic / screening or monitoring results. Similarly, researchers may only be given aggregated data that is not personally identifiable to avoid violating privacy laws and regulations.

[0242] The IO 2041 provides connections to the clients 260a-n as well as to the user data 220 and the model server(s) 230. Collected speech and video data (in some embodiments raw audio and video files) is provided by the IO 2041 to the user data 220, the runtime model server(s) 2010, and the training data filter 2001. The sign data from the clients 260a-n is provided to a sign dataset 2021 in the user data 220. This may be stored in various databases 2023. The sign data includes verified diagnosed patients as well as putative signs collected from specific user attributes or human annotations. Client ID information and logs may be provided to the user data 220 from the IO 2041 as well. The user data 220 may be further enriched with clinical and social records 210 sourced from any number of third party feeds. This may include social media information obtained from web crawlers, EHR databases from medical professionals, public health data sources, etc.

[0243] The training data filter 2001 may consume speech and video data and append labeling data 2021 to generate a training data set. This training data set is provided to model training server(s) 2030 for generation of a set of machine learning models. The models are stored in a model repository 2050 and utilized by runtime model server(s) 2010 to make screening or monitoring outcome determinations in addition to generating other descriptors for clients 260a-n. The model repository 2050 together with the model training server(s) 2030 and the runtime model server(s) 2010 constitute the model server(s) 250. The runtime model server(s) 2010 and the model training server(s) 2030 are described in more detail below in connection with FIG. 20B and FIG. 21, respectively.

[0244] In FIG. 20B, the runtime model server(s) 2010 are provided in more detail. The server receives speech and video input originating from clients 260a-n. A signal preprocessor multiplexer 2011 performs conditioning on the input data, such as removing noise or other artifacts in the signal that may cause modeling errors. These signal processing and data preparation tasks include diarization, segmentation, and noise reduction on both the speech and video signals. Additionally, metadata may be layered onto the speech and video data. This data may be provided in this preprocessed form to a bus 2014 for consumption by a modeler 2020, or may be subjected to any number of third-party off-the-shelf automatic speech recognition (ASR) systems 2012. The output of the ASR 2012 includes a machine-readable transcription of the spoken portion of the audio data. This ASR 2012 output is in turn provided to a bus 2014 for consumption by subsequent components. The signal pre-processor multiplexer 2011 may be provided with confidence values ​​such as audio quality (signal quality, length of samples) and transcription confidence (how accurate the transcription is) values ​​2090,2091.

[0245] 20B also includes a metadata model 2018. The metadata model may analyze patient data, such as demographic data, medical history data, and patient-provided data.

[0246] Additionally, clinical, demographic, and social data may be presented to the bus 2014 for subsequent use by the modeler 202. Finally, a model reader 2013 may access the secured models from a model repository 2050, which is also provided to the bus 2014. The modeler 2020 consumes the models, pre-processed audio and visual data, and the output of the ASR 2012 to analyze the responses of the clients 260a-n with respect to the subject's health condition. Unlike previous systems for monitoring health conditions, the current system includes a natural language processing (NLP) model 2015, an acoustic model 2016, and a video model 2017, all of which work in concert to generate a classification of the health condition of the clients 260a-n. These modelers not only work in concert, but also consume each other's output to refine the model output. Each modeler and the manner in which they adjust to improve their classification accuracy will be described in more detail in conjunction with subsequent figures.

[0247] The output of each of these modelers 2020 is provided separately to a calibration, confidence, and desired descriptor module 2092, which calibrates the output to generate a scaled score and provides a confidence measure to the score. The desired descriptor module may assign a human readable label to the score. The output of the desired descriptor module 2092 is provided to a model weights and fusion engine 2019, which combines the model outputs into a single unified classification of the health status of each client 260a-n. Model weighting may be performed using static weights, such as weighting the output of the NLP model 2015 over the output of either the acoustic model 2016 or the video model 2017. However, more robust dynamic weighting methods may be applied as well. For example, the weight on a given model output may be changed in some embodiments based on the confidence level of the model's classification. For example, if the NLP model 2015 classifies an individual as not depressed with a confidence of 0.56 (out of 0.00 to 1.00), but the acoustic model 2016 classifies an individual as depressed with a confidence of 0.97, then sometimes the weights of the model outputs may be weighted such that a greater weight is given to the acoustic model 2016. In some embodiments, the weight of a given model may be linearly scaled by the confidence level multiplied by the base weight for the model. In yet other embodiments, the model output weights are based on time. For example, the NLP model 2015 may generally be given a greater weight than other models, but if the user is not speaking, the video model 2017 may be given a greater weight across that time domain. Similarly, if the video model 2017 and the acoustic model 2016 independently suggest that the person is nervous and lying (frequent gaze shifts, increased sweating, upward pitch modulation, increased speech rate, etc.), then the weight of the NLP model 2015 may be minimized since it is likely that the individual is not answering questions truthfully.

[0248] After fusion and weighting of model outputs, the resulting classifications may be combined with features and other user information in a multiple output module 2051 to generate final results. As previously mentioned, these results are provided to the original user data 220 for storage and potentially future training material, as well as to the results and presentation module 2045 of the web server 240 for display, at least in part, to the clients 260a-n and stakeholders 270a-n. These results are in turn used by the dialogue engine 2043 to adapt the dialogue with the clients 260a-n.

[0249] Referring now to FIG. 21, the model training server(s) 2030 are provided in more detail. Similar to the runtime model server(s) 2010, the model training server(s) 2030 consume a set of data sources. However, these data sources have been filtered by a training data filter 2001 to provide only data regarding which label information is known or attributable. The model training server takes as input a voice quality confidence value 2095 (which may include bit rate, noise, and duration of the audio signal) and a transcription confidence value 2096. These confidence values ​​may include the same types of data as in FIG. 20B. The filtered social, demographic, and clinical data, audio and video data, and label data are all provided to a pre-processor 2031 for cleaning and normalization of the filtered data sources. The processed data is then provided to a bus 2040 for consumption by various trainers 2039 and also to one or more third party ASR systems 2032 for generation of ASR outputs that are also provided to the bus 2040. The signal preprocessor multiplexer 2011 may be provided with confidence values ​​such as voice quality (signal quality, length of samples) and transcription confidence (how accurate the transcription is) values ​​2095, 2096.

[0250] The model trainer 2039 consumes the processed audio, visual, metadata, and ASR output data in the NLP trainer 2033, the acoustic trainer 2034, the video trainer 2035, and the metadata trainer 2036. The trained models are individually provided to a calibration, confidence, and desired descriptors module 2097, which calibrates the output to generate a scaled score and provides a confidence measure to the score. The desired descriptors module may assign a human readable label to the score. The trained and calibrated models are provided to a fusion model trainer 2037 to combine the trained models into a trained combined model. The individual and combined models may be stored in a model repository 2050. Additionally, the trained models may be provided to a personalizer 2038, which leverages metadata (such as demographic information and data collated from social media streams) to specifically tailor the models to a given client 260a-n.

[0251] For example, a specific model xo may be generated to classify an acoustic signal as representing either someone who is depressed or someone who is not depressed. The tenor, pitch, and cadence of the speech input may vary significantly between young and older individuals. Therefore, a specific model is developed based on whether the patient being screened is young or older (models xy and xe, respectively). Similarly, women generally have more variability in their acoustic signals compared to men, suggesting that a further set of acoustic models is needed (models xf and xm, respectively). It is also clear that combination models are desirable for young vs. older women and young vs. older men (models xyf, xef, xym, and xem, respectively). Clearly, as more personalized groupings are generated, the possible number of applicable models grows exponentially.

[0252] In some embodiments, where metadata about an individual provides insight into their age, gender, ethnicity, educational background, accent / region they grew up in, etc., this information may be utilized to select the most appropriate model to use in future interactions with this given patient, and this information may also be used to train models that are applied to individuals sharing similar attributes.

[0253] In addition to personalizing models based on population segments and attributes, the personalizer 2038 may also personalize a model or set of models for a particular individual based on that individual's past history and the indicator data known about that individual. This activity is more computationally expensive than relying on modeling the entire population or segment, but produces more accurate and granular results. All personalized models are provided from the personalizer 2038 to the model repository 2050 to be kept until needed for patient evaluation.

[0254] Then, during analysis, the client 260a-n may be identified first and, if possible, a personalized model may be used for screening or monitoring the client. If not available, but metadata is known about the individual, the most specific model for the most specific segment is used for screening or monitoring. If no metadata is available, the model selected is the general population-wide model. Utilizing such a tiered modeling structure, the more information known about the client 260a-n, the more specific and accurate models can be employed. Thus, for each client 260a-n, the "best" model is utilized given the data available for that client.

[0255] A general overall flow of information is shown in (FIG. 22). The assessment test administrator 2202 of the real-time system 302 has an interactive conversation with the patient via the patient device 312. The patient's responsive audiovisual signals are received by the real-time system 302 from the patient device 312. The exchange of information between the real-time system 302 and the patient device 312 may be via a dedicated app running on the patient device 112 or via a traditional video call between the patient device 312 and the video call logic of the assessment test administrator 2202. It should be understood that while this exemplary embodiment uses audiovisual signals to evaluate the patient's condition, in another embodiment, good results can be obtained using audio-only signals. In such another embodiment, a normal audio-only telephone conversation can serve as a means for evaluation by the assessment test administrator 2202.

[0256] In a manner described more fully below, the assessment test administrator 2202 uses the composite model 2204 to assess the patient's condition in real time, i.e., as the verbal conversation occurs. Such intermediate assessments are used, in a manner described more fully below, to control the conversation to make it more responsive and therefore more engaging for the patient, and to help keep the conversation as short as possible while maintaining the accuracy of the final assessment.

[0257] The modeling system 304 receives collected patient data 2206, including audiovisual signals of the patient during the assessment test. In embodiments where the assessment test involves a patient device 312, a dedicated app running on the patient device 312, the modeling system 104 may receive collected patient data 2206 from the patient device 312. Alternatively, in embodiments where the assessment test involves a video or voice call with the patient device 312, the modeling system 304 receives collected patient data 2206 from the real-time system 302.

[0258] The modeling system 304 retrieves clinical data 2220 from the clinical data server 306. The clinical data 2220 generally includes any available clinical data relating to patients, other patients evaluated by the assessment test administrator 2202, and the general public that may be useful in training any of the various models described herein.

[0259] Pre-processing 2208 prepares any audiovisual data for optimal analysis. Having a high quality signal to start with is very helpful in performing accurate analysis. Pre-processing 2208 is shown within the modeling system 304. In another embodiment, pre-processing is included in the real-time system 302 to improve accuracy in the application of the composite model 204.

[0260] Speech recognition 2210 processes the speech represented in the audiovisual data after preprocessing 2208, including automatic speech recognition (ASR). The ASR may be conventional. Language model training 2212 uses the results of speech recognition 2210 to train a language model 214.

[0261] Acoustic model training 2216 trains an acoustic model 2218 using the audiovisual data after preprocessing 2208. Visual model training 2224 trains a visual model 2226 using the audiovisual data after preprocessing 2208. To the extent sufficient data (both collected patient data 2206 and clinical data 2222) is available for the target patient, language model training 2212, acoustic model training 2216, and visual model training 2224 train the language model 2214, acoustic model 2218, and visual model 2226, respectively, specifically for the target patient. The training may also use clinical data 2222 for patients who share one or more phenotypes with the target patient.

[0262] In a manner described more fully below, composite model builder 2222 uses language model 2214, acoustic model 2218, and visual model 2226 in combination with clinical data 2220 to combine the language model, acoustic model, and visual model into composite model 2204. Assessment test administrator 2202 then uses composite model 2204 in real time to assess the subject patient's current condition and respond accordingly with verbal dialogue to the subject patient, as described more fully below.

[0263] As previously described, the assessment test administrator 2202 administers the depression assessment test to the subject patient by engaging in an interactive verbal conversation with the subject patient via the patient device 312 .

[0264] Attention now turns to the particular models used by the runtime model server(s) 2010. Turning to Figure 23A, a general block diagram of one example instantiation of the acoustic model 2016 is provided. The speech and video data are provided to a high level feature renderer 2320 which operates in concert with a temporal dynamics modeler 2330. Influencing the operation of these components is a model conditioner 2340 which consumes features from the descriptive features 2018, results generated from the speech and video models 2015, 2017, respectively, as well as clinical and social data.

[0265] Returning to the acoustic model 2016, the high level feature representer 2320 and the temporal dynamics modeler 2330 also receive the output of the raw higher level feature extractor 2310, which identifies features in the incoming acoustic signal and feeds them to the model. The high level feature representer 2320 and the temporal dynamics modeler 2330 generate acoustic model results that may be fused into a final result that classifies the health status of an individual, or may be consumed by other models for tuning purposes.

[0266] The high level feature indicator 2320 includes existing models that leverage frequency, pitch, amplitude, and other acoustic features that provide valuable insight into feature classification. Several off-the-shelf "black box" algorithms accept audio signal inputs and provide concomitant accuracy in classifying emotional states. For example, emotions such as sadness, happiness, anger, and surprise can already be identified in audio samples using existing solutions. Further emotions such as jealousy, nervousness, excitement, merriment, fear, disgust, trust, and anticipation can also be leveraged as they progress. However, the current system and method goes further by matching these emotions, their intensity, and confidence in the emotions to patterns in emotional profiles that are indicative of specific mental health conditions. For example, pattern recognition may be trained to identify respondent emotional states that indicate depression based on patients known to suffer from depression.

[0267] FIG. 23B illustrates an embodiment of FIG. 23A including an acoustic modeling block 2341. The acoustic modeling block 2341 includes several acoustic models. The acoustic models may be separate models using machine learning algorithms. The exemplary list of models shown in FIG. 23B is not necessarily a comprehensive list of possible models. These models may include a combination of existing third-party models and internally derived models. FIG. 23B includes an acoustic embedding model 2342, a spectrotemporal model 2343, an acoustic effects model 2345, a speaker personality model 2346, an intonation model 2347, a time / speaking rate model 2348, a pronunciation model 2349, and a fluency model 2361. The machine learning algorithms used by these models include neural networks, deep neural networks, support vector machines, decision trees, hidden Markov models, Gaussian mixture models, etc.

[0268] 23C illustrates the score calibration and confidence module 2370. The score calibration and confidence module 2370 includes a score calibration module 2371 and a performance estimation module 2374. The score calibration module 2371 includes a classification module 2372 and a mapping module 2373.

[0269] The score calibration and confidence module 2370 may accept as input raw scores generated by a machine learning algorithm, such as a neural network or deep learning network, that may be analyzing the audiovisual data. The score calibration and confidence module 2370 may also accept a set of labels on which the data should be classified. The labels may be provided by a clinician. The classification module 2371 may apply one or more labels to the raw score based on the value of the score. For example, if the score has a probability close to 1, the classification module 2371 may apply a "severe" label to the score. The classification module 2371 may apply the labels based on criteria set by the clinician or may algorithmically determine the labels for the score, for example using a machine learning algorithm. The mapping module 2372 may scale the raw scores to fall within a numerical range, such as 120-180 or 0-700. The classification module 2371 may operate before or after the mapping module 2372.

[0270] After calibrating the data, the score calibration and confidence module 2370 may determine a confidence index 2376 by estimating performance for the labeled and scaled scores. The performance may be estimated by analyzing features of the collected data, such as duration, timbre, accent, and other features. The estimated performance may be a weighting parameter applied to the score. This weighting parameter may form the score confidence.

[0271] To provide more detailed context and explanation regarding the operation of the acoustic model 2016, a highly simplified single instantiation of one possible version of the high-level feature indicator 2320 is provided with reference to Figure 24. It should be noted that this example is provided for illustrative purposes only and is not intended to limit the embodiment of the high-level feature indicator 2320 in any way.

[0272] In this exemplary embodiment, the raw high level feature extractor 2310 takes the acoustic data signal and converts it into a spectrogram image 2321. Figure 55 gives an example of an image of such a spectrogram 5500 of a person speaking. This kind of spectrogram gives information about the audio signal frequency, the signal amplitude (presented here in terms of intensity / frequency labeling darkness) and time along one axis. Such a spectrogram 5500 is considered as the raw features of the acoustic signal, like pitch, cadence, energy level, etc.

[0273] A spectrogram sampler 2323 then selects a portion of the image over a fixed time frame, e.g., time 0-10 seconds is one standard sample size, although other sample time lengths are contemplated. Figure 56 provides an example of a sampled portion 5502 of a spectrogram 5600. This image data is then represented, in this particular non-limiting example, as an MxN matrix (x). Equations that include x as a variable and have a known solution are then processed to determine estimates of the unknown variables (matrices and vectors) in the equations. For example,

number

[0274] A linear equation such as: may be utilized. As mentioned above, the solution y is known.

[0275] This is because the unknown variable (w in this example equation) Tand b) determining a set of randomized guesses. The equations are solved using these guessed variables and the error of the solved solution is calculated using the known solution values. The error is calculated as follows:

number

[0276] By repeating this process thousands, if not millions, of times, values ​​of the variables that approximate the actual variable values ​​may be determined. This is brute force regression, in which case the error values ​​(

number

[0277] This approximation is an abstraction of the mental state being tested according to the equation entered. The system may have pre-determined thresholds or cut-off values ​​2322 for variables that indicate whether a response is indicative of a mental state or not. These cut-off values ​​are trained by analyzing responses from individuals whose mental states are already known.

[0278] Determining the equations may utilize deep learning techniques, as described above. This may include recurrent neural networks 2324 and / or convolutional neural networks 2325. In some cases, for example, long short-term memories (LSTMs) or gated recurrent units (GRUs) may be used. In this way, depression or alternative mental states may be analyzed directly in the acoustic portion of the response. This, in combination with the use of an off-the-shelf emotion detection "black box" system with pattern recognition, may result in a more robust classification by the mental state classifier 2326 based on the acoustic signal, which in this example is provided as the acoustic analysis output 2327.

[0279] As mentioned above, this example of using a spectrogram as a feature of the analysis is only one of many possible instantiations of the activity of the high level feature indicator 2320. Other features and mechanisms for processing these features may be analyzed as well. For example, pitch level, isolated breathing patterns, total energy of the audio signal, etc., may all be subject to similar time-based analysis to classify the features as indicative of health conditions.

[0280] 25, the NLP model 2015 is provided in more detail. The system consumes output from the ASR system 2012 and performs post-processing thereon by the ASR output post-processor 2510. This post-processing includes adjustments to the ASR output (when there are multiple outputs). Post-processing may also include n-gram generation, parsing activities, etc.

[0281] Similarly, results from the video and acoustic models 2016, 2017, respectively, as well as clinical and social data, are consumed by the model conditioner 2540 to modify the functioning of the language model 2550. The language model 2550 works in concert with the temporal dynamics modeler 2520 to generate the NLP model results.

[0282] The language model 2550 includes several separate models. The exemplary list of models shown in FIG. 25 is not necessarily a comprehensive list of possible models. These models may include a combination of existing third-party models and internally derived models. The language models may use standard machine learning or deep learning algorithms as well as language modeling algorithms such as n-gram. For example, the emotion model 2551 is a readily available third-party model that uses either the original text samples or verbal samples that have been transcribed by a human or machine speech recognizer to output a determination of whether the emotion of the discussion is generally positive or negative. In general, positive emotion is inversely correlated with depression, while negative emotion is correlated with depression classification.

[0283] The statistical language model 2552 utilizes n-gram and pattern recognition in the ASR output to statistically match patterns and n-gram frequencies with known indicators of depression. For example, a particular set of words may be statistically indicative of depression. Similarly, a particular vocabulary and word types used by a speaker may be indicative of depression or the absence of depression.

[0284] The topic model 2553 identifies the types of topics in the ASR output. Certain topics such as death, suicide, despair, and worthiness (or lack thereof) may all be positively correlated with a classification of depression. Furthermore, there is a potential negative correlation between behavior (indicated by verb usage) and depression. Thus, an ASR output with high verb usage may indicate that the client 260a-n is not depressed. Furthermore, topic modeling based on known questions or prompts given to the subject can produce better performance by using topic-specific models that are pre-trained to process answers related to mental health conditions.

[0285] The syntax model 2554 identifies situations where the focus of the ASR output is internal versus external. The use of terms such as "I" and "me" indicates an internal focus, while terms such as "you" and "they" indicate less internal focus. More internal focus has been identified as generally correlated with increased chances of depression. The syntax model 2554 may also look at the complexity of the utterance. Depressed individuals tend to have reduced sentence complexity. Furthermore, energy levels, indicated by strong or polarized language, are negatively correlated with depression. Thus, a person with very simple, internally focused sentences and low energy descriptive words indicates a classification of depression.

[0286] The embedding and clustering model 2556 may map words to prototypical words or word categories. For example, the terms "kitten," "cat," and "kitty" may all be mapped to the term "cat." Unlike other models, the embedding and clustering model 2556 does not generate a direct indication of whether a patient is depressed or not; rather, the output of this model is consumed by other language models 2550.

[0287] The dialogue and discourse model 2557 specifies the latency and usage of spacer words (such as "like", "umm") Additionally, the dialogue and discourse model 2557 specifies dialogue acts such as questions versus utterances.

[0288] Affect or emotion model 2558 provides scores, generally posterior probabilities, for a set of predefined emotions (e.g., happy, sad) that represent how well a sample matches a pre-trained model for each said emotion. These probabilities may then be used in various ways as inputs to mental health models and / or in a transfer learning setting. Speaker personality model 2559 provides scores, generally posterior probabilities, for a set of predefined speaker personality traits (e.g., pleasantness, openness) that represent how well a sample matches a pre-trained model for each said trait. These probabilities may then be used in various ways as inputs to mental health models and / or in a transfer learning setting.

[0289] The non-verbal model using ASR events 2561 may provide a score based on the patient's non-vocabulary utterances that may be indicative of a mental state in any way. These utterances may be laughs, sighs, or deep breaths that may be picked up and transcribed by the ASR.

[0290] The text quality confidence module 2560 determines a confidence index for the output of the ASR output post-processor 2510. The confidence index may be determined based on text metadata (demographic information about the patient, environmental conditions, recording method, etc.) as well as context (e.g., length of the speech sample, question asked).

[0291] It should be noted that each of these models influences the other and may affect the results and / or how these results are classified. For example, low energy verbal responses are generally indicative of depression, whereas high energy wording is negatively correlated with depression.

[0292] Turning now to the video model 2017 of Figure 26, we again refer to a group of feature extractors 2610 that consume video data. Within the feature extractor 2610 is a face bounder 2611 that recognizes the edges of a person's face and extracts this region of the image for processing. Clearly, facial features provide significant input into how an individual is feeling. Sadness, fatigue, and worry are all associated with a depressed state, whereas joy, excitement, and positivity are all negatively correlated with depression.

[0293] Furthermore, more specific boundaries may be considered, for example, the area around the eyes may be analyzed separately from the area around the mouth. This allows for more emphasis on different image regions based on context. In this set of examples, the area around the mouth generally gives a lot of information about the mood of an individual, but when a person is speaking, this data is likely to be inaccurate due to the movements associated with speech formation. Acoustic and language models may provide insight into when a user is speaking to reduce the reliance on the analysis of the mouth region extraction. In contrast, the area around the eyes is generally very expressive when someone is speaking, so the reliance on this feature is relied on more when the individual is speaking.

[0294] The posture tracker 2612 can see or see larger body movements or positions. A crouched posture indicates anxiety, sadness, and other characteristics indicative of depression. The presence of excessive fidgety behavior or conversely abnormal stillness is also indicative of depression. However, moderate movements and fidgety behavior are not associated with depression. Similarly, upright posture and relaxed movements are inversely associated with depression classification. Finally, even the orientation in which an individual sits or stands is indicative of depression. A user who faces the camera directly is less likely to be depressed. In contrast, an individual who positions their body at an angle to the camera or who otherwise covers themselves (e.g., by folding their arms) is more likely to be depressed.

[0295] The gaze tracker 2613 is particularly useful for determining where a user is looking and when a person's gaze shifts (in response to what stimuli). Looking at the screen or camera of the client device 260a-n indicates engagement, trustworthiness, honesty - all of which are signs of a non-depressed state. On the other hand, constantly looking down is suggestive of depression. A constantly changing gaze indicates nervousness and dishonesty. Such feedback may be used by the NLP model 2015 to reduce the value of analysis based on semantics during this period, as the individual is more likely to be evasive in their answers and / or outright lying. This is especially true when gaze patterns change dramatically in response to stimuli. For example, if the system asks an individual if they have thoughts of self-harming, and suddenly the user looks away from the camera and the gaze shifts, the denial of such thoughts (which is traditionally considered an indication of a positively non-depressed state) is ignored. Rather, emphasis is placed on the output of the acoustic model 2016 and video model 2017 from the analysis of other extracted features.

[0296] The image processing feature extractor 2614 may take the form of any number of specific feature extractions, such as emotion identifiers, speech identifiers (from video as opposed to audio data), and the specific boundary extractors disclosed above (e.g., areas around the eyes). All of the extracted features are provided to a high level feature displayer 2620 and a classifier and / or regressor 2630, which work in conjunction to generate the video model results. As with the other models, the video model 2017 is influenced by the outputs of the NLP model 2015 and acoustic model 2016, as well as clinical and social data. The model conditioner 2640 uses this information to modify what analysis is performed or the weighting given to any particular findings.

[0297] The descriptive features module 2018 of FIG. 27 includes direct measurements 2710 and model outputs 2720 resulting from analysis of speech and video data. The descriptive features module may not be included in either the runtime model server 2010 or the model training server 2030. Instead, the descriptive features may be incorporated into the acoustic and NLP models. Disclosed in the description of FIG. 27 are examples of descriptive features. Many different measurements 2710 and model outputs 2720 are collected by the descriptive features 2018 module. For example, measurements include a speaking rate analyzer 2711 that tracks a speaker's words at least every minute. Faster speech generally indicates excitement, energy, and / or nervousness.

[0298] On the other hand, a slow speech rate indicates the presence of hesitation, lethargy, or a difficult topic. Alone, this measure has little value, but when used as an input for other models, the speech rate provides a context that allows for more accurate classification by those other models. Similarly, the energy analyzer 2713 measures the total acoustic energy of the speech components. An increase in energy may indicate emphasis on a particular part of the dialogue, a general excitement or lethargy level, etc. Again, such information alone is of little use in determining whether a person is depressed, but in combination with other models, it helps ensure that an appropriate classification is made. For example, if the energy level increases when a person talks about their pet dog, the system determines that this topic is of interest to the individual, and the dialogue may be steered towards this topic if a longer dialogue is needed to gather more user data for analysis. The time analyzer 2715 determines the date, time, week, and year to provide context over the dialogue. For example, people are generally more depressed in the winter months, around certain holidays, and on certain days and times of the week. All this timing information can be used to alter the dialogue (by providing topicality) or by allowing classification thresholds to be slightly altered to reflect these trends.

[0299] The model output 2720 may include a topic analyzer 2721, various emotion analyzers 2723 (e.g., anxiety, happiness, sadness), a sentiment analyzer 2725, an engagement analyzer 2727, and an arousal analyzer 2729. Some of these analyzers may work in other models as well, for example, the NLP model 2015 already includes an emotion model 2551, but the sentiment analyzer 2725 in the descriptive features 2018 module works independently from other models and includes different input variables even if the output is similar.

[0300] The engagement analyzer 2727 operates to determine how engaged the client 260a-n is in the interaction. A high level of engagement tends to indicate sincerity and enthusiasm. The arousal analyzer 2729 gives insight into how energetic or lethargic the user is. An important feature of the descriptive features 2018 module is that each of these features, whether measured or the result of a model output, is normalized by the individual by the normalizer 2730. For example, some people simply speak faster than others, and a higher measurement of words per minute for this individual relative to others may not indicate anything abnormal. The degree of any of these features is adjusted by the normalizer 2730 according to the baseline level of the particular individual. Obviously, the normalizer 2730 operates more accurately the more data that is collected about any given individual. A first interaction with a client 260a-n may not be immediately and effectively normalized, but as the interaction progresses, the ability to determine a baseline for this person's speech rate, energy level, engagement, general mood / attitude, etc. can be more easily ascertained using standard statistical analysis of changes in these characteristics over time. This becomes especially true after multiple interactions with any given individual.

[0301] After normalization, the system may identify trends in these features for the individual through analysis by trend tracker 2740. The trend tracker divides the interactions by time domain and looks for changes in values ​​between various time periods. Statistically significant changes, especially those that persist over multiple time periods, are identified as trends in the features for this individual. All features, both in raw and normalized form, as well as any trends, are output as descriptive results.

[0302] Although not addressed in any of the figures, it is fully within the scope of embodiments of the present disclosure that additional models may be used to make classifications regarding the health status of clients 260a-n using alternative data sources. For example, it has been discussed that a client device may be able to collect biometric data (temperature, skin chemistry data, pulse rate, motion data, etc.) from an individual during an interaction. Models focused on these inputs may be leveraged by runtime model server(s) 2010 to arrive at decisions based on this data. The disclosed system may identify chemical markers in skin (e.g., cortisol), sweat, temperature changes (e.g., flushing), heart rate changes, etc., for diagnostic purposes.

[0303] Having discussed in some depth the details of the runtime model server(s) 2010, attention is now directed to the dialogue engine 2043, as will be seen in more detail in connection with FIG. 28. A process flow diagram characterizing the components of the dialogue engine 2043 is depicted in FIG. 8. The dialogue engine 2043 dictates the dialogue between the web server(s) 240 and the clients 260a-n. As previously mentioned, these dialogues may consist of question and answer sessions, with set numbers and sequences or questions. In such an embodiment, this type of assessment is substantially an automated version of those previously utilized for depression diagnosis, except for audio and video capture for screening or monitoring accuracy improvement. Such questions and answers may be conducted with text questions displayed on the client device or by verbal recording of the questions. However, such systems are generally not particularly involved with the clients 260a-n, and may allow the dialogue to be completed in good faith or the dialogue to be terminated early. It is therefore desirable to have dynamic dialogues, which requires a more sophisticated dialogue engine 2043 as seen in this figure.

[0304] The dialogue engine 2043 includes the ability to take several actions, such as different prompts, questions, and other dialogues. These are stored in a question and action bank 2810. The dialogue engine 2043 also includes a history and state machine 2820 that keeps track of what has already happened in the dialogue and the current state of the dialogue.

[0305] The status and history information, database of possible questions and actions, and further data are consumed by the interaction modeler 2830 to determine the next step in the interaction. Other information consumed consists of user data, clinical and social data with which the client is interacting, as well as model results, NLP output, and descriptive feature results. The user data, clinical data, and social media data are all consumed by the user preference analyzer 2832 to reveal the user's preferences. As mentioned above, appealing to the user is one of the major hurdles to successful screening or monitoring. If the user does not want to use the system, they may not use the system in the first place or may end the interaction prematurely. Alternatively, an unpleasant interaction may make the user less honest and open about the system. The inability to properly screen an individual for depression or health conditions is a serious problem in general. This is because these individuals are more likely to continue to suffer from their illness without assistance, or even worse, die prematurely. Thus, a high degree of user engagement can literally save lives.

[0306] By determining preference information, the interaction is tailored in a manner that appeals to the user's interests and desires. Topics identified in social media feeds are incorporated into the interaction to pique the user's interest. Preference data collected from the user tailors the interaction to be more user-friendly, and the user's specific needs or limitations revealed in the clinical data are also utilized to make the interaction experience user-friendly. For example, if the clinical data includes information that the user experiences hearing loss, the amount of interaction may be increased proportionately to make the interaction easier. Similarly, if the user indicates that their preferred language is Spanish, the system may automatically conduct the interaction in this language.

[0307] In response, the descriptive features and model results are used by the user response analyzer 2831 to determine whether the user has answered a question (if the interaction is in question-answer format) or, once enough data has been collected to provide an appropriate classification, whether the interaction is more of a "free-form" conversation or even a monologue by the client on a topic of interest.

[0308] Additionally, a navigation module 2834 receives the NLP output and semantically analyzes the NLP results for command language in near real-time. Such commands may include utterances such as "Can you repeat that?", "Speak up," "I don't want to talk about that," etc. These types of "command" phrases indicate to the system that an immediate action is requested by the user.

[0309] The output from each of the navigation module 2834, the user response analyzer 2831, and the user preference analyzer 2832 are provided to the action generator 2833, along with access to the question and adaptive action bank 2810 and the history and state machine 2820. The action generator 2833 applies a rule-based model to determine which action in the question and adaptive action bank 2810 is appropriate. Alternatively, a machine learning model is applied instead of a rule-based decision model. This results in the output of a customized action that is provided to the IO 2041 for communication to the client 260a-n. The customized action is similarly sent back to the history and state machine 2820 so that the current state and past actions are appropriately recorded. The customized action may include, for example, asking a specific question, prompting a topic, switching to another voice or language, terminating a dialogue, changing the volume of the dialogue, changing the speech rate, font size, and color, etc.

[0310] Having described the structure and system of the health screening or monitoring system 2000 in considerable detail, attention is now directed to one example process 2900 of health screening or monitoring of a client. In this example process, clinical and social data about the client is collated and stored in a data store (at step 2910). This information may be collected from social media platforms utilizing crawlers or similar means. Clinical data may be collected from health networks, physicians, insurance companies, and the like. In some embodiments, the health screening or monitoring system 2000 may be deployed as an extension of the care provider, which may allow such sharing of clinical data with reduced concerns regarding violation of privacy laws (such as HIPAA). However, if the health screening or monitoring system 2000 is operated as a separate entity outside of the healthcare network, further consents, encryption protocols, and removal of personally identifiable information may be required to allow open sharing of clinical data while remaining compliant with applicable regulations. Clinical data may include electronic health records, physician notes, medications, diagnoses, and the like.

[0311] The process may then require that a model be available to analyze the client's interactions. An initial dataset including the labeled data (definitive or imputed diagnoses of depression) is fed to a series of trainers, which train individual models and then fuse them into a combined model (at 2920). Such training may include personalizing the models if further metadata is available.

[0312] Figure 30 provides a more detailed illustration of one example of a process for such model training. As previously described, indicator data is received (at 3010). The indicators include a definitive diagnosis of depression (or other health condition being screened for). Similarly, circumstances in which the indicators may be imputed or otherwise inferred are used to augment the training data set.

[0313] The attributable indicator data is received by manual review of medical records and / or interaction records with a given client. For example, in a predictive mode, if an indicator is unknown, it can be determined whether an indicator can be inferred for a data point given other information such as patient records, system predictions, clinically validated surveys or questionnaires, and other clinical data. Due to the relative rarity of indicator datasets and the need for a large number of training samples to generate accurate models, it is often important that the indicator data include not only confirmed cases of depression but also these putative indicators.

[0314] The process further includes receiving (at 3020) filtered data. This data is filtered such that only data for which the indicators are known (or estimated) is used. Next, each model is trained. Such training includes training an NLP model (at 3030), an acoustic model (at 3040), a video model (at 3050), and descriptive features (at 3060). It should be noted that these training processes may be performed in any order or trained in parallel. In some embodiments, the parallel training includes generating interdependencies between the various models. These interdependencies are one of the key features that uniquely enable the systems and methods disclosed herein to provide improved and highly accurate classification of health conditions.

[0315] The resulting trained models may be fused or aggregated, and the final fused trained model may be stored (at 3070). The models (both individual and fused) are stored in a model repository. However, it may also be desirable to generate (at 3080) model variants that are customized for different population groups or even specific individuals.

[0316] The process for customizing and personalizing the model is described in more detail in relation to Figure 31. Personalization relies on metadata that is stored with the filtered training data. This metadata is received (at 3081). Specific population segment features are identified in the metadata and extracted (at 3082). These segment features are used to train a model specific to that segment. This is achieved by clustering (at 3083) the filtered training data by these segmentation features. A given training piece may be included in several possible segments, each of which may be non-overlapping or of successively increasing granularity.

[0317] For example, assume that labeled training data is received from a known individual. In this example, the individual is identified as a black woman in her 70s. This training data is then used to train on models specific to African American individuals, African American women, women, seniors, elderly women, African American seniors, and African American elderly women. Thus, this single piece of training data is used to generate seven different models, each with slightly different scopes and granularity levels. In situations where age is further divided, this number of models to be trained from this data is further increased (e.g., adult women, women over 50, women over 70, individuals over 70, etc.). Models are then trained for each of these segments (at 3084). The customized models are annotated (at 3085) with the segment(s) to which they can be applied, allowing for easy retrieval of customized models when a new response is received with respect to a classification where information about the individual is known and this response can be utilized to select the model most appropriate / tailored for this person.

[0318] This is important because in many cases, a model for identifying health conditions for one individual may be completely inappropriate for classifying another. For example, a Caucasian individual may require a different video model compared to a person of African descent. Similarly, men and women often have different acoustic characteristics that require different acoustic models to be leveraged to accurately classify them. Even a woman in her early 20s sounds different from a woman in her 50s, who also sounds different from a woman in her 80s. NLP models for native speakers versus second language speakers may differ drastically as well. Even between generations, NLP models differ drastically to address differences in slang and other speech nuances. Making models available to individuals at various levels of granularity greatly improves the accuracy of classification by these models, as they can apply the most appropriate model.

[0319] Returning to FIG. 30, after this personalization is complete, the customized model is also stored (at 3090) in the model repository along with the original model and the fused model. It should be noted that while model customization generally improves classification accuracy, any such accuracy improvement is jeopardized if the number of training datasets available for the model is low. The system tracks the number of training datasets used to train any given customized model, and only models with sufficiently large enough training sets are marked as "active" in the model repository. Active models can be used by the runtime model server(s) 2010 to process newly received response data. Inactive models are only stored until enough data has been collected to properly train them, at which point they are updated as active.

[0320] Returning to Figure 29, after model training, the process may involve interaction with a client (at 2930). This interaction may consist of a question and answer style format, a free-flowing conversation, or even the client providing topic prompts and monologue style input.

[0321] Figure 32 shows an example of this dialogue process. First, the system needs to know the current state of the dialogue (at 3210) and the historical actions taken in the dialogue. A state machine and a log of previous actions provide this context. The process also receives user, clinical, and social data (at 3220). This data is used to extract user preference information (at 3230). For example, preferences may be explicitly indicated in the user data, such as language preferences, topics of interest, etc. Alternatively, these preferences are extracted from clinical and social data. For example, social data provides a wealth of information about topics that a user is interested in, and clinical data provides insight into accessibility issues, etc.

[0322] Additionally, model results are received (at 3240), which are used to analyze the user's responses (at 3250) and make decisions regarding the validity of the data already collected. For example, if the model results determine that there is still no clear classification, the dialogue may focus on collecting more data in the future. Alternatively, if enough data has been collected to make a reliable classification, the dialogue may instead focus on a solution. Additionally, the dialogue management may receive direct command utterances / navigation commands from the user (at 3260). This may include actions such as repeating the last dialogue exchange, increasing or decreasing the volume, rephrasing the question, requesting more time, requesting to skip a topic, etc.

[0323] All this information is consumed by the action generator (at 3270) to determine the best course of action for the next step. An action is selected from the question and adaptive action bank depending on the current state (and previous history of the interaction) and any commands, preferences, and results already received. This can be completed using a rules-based engine in some embodiments. For example, a direct navigation command may be preferred over another action, but in the absence of a command uttered by the user, the model response can be matched against the current state to determine if the goal of the state has been achieved. If so, an action is selected from the repository that satisfies another goal that did not occur in the interaction history. This action is also modified based on preferences, if possible. Alternatively, the action selection is based on a machine learning model (rather than a rules-based system).

[0324] The customized actions are used to manage the interaction with the client and also to update the current and past state activities (at 3280). The process checks (at 3290) whether the goal is met and whether the interaction should be terminated. If not, the entire process can be repeated with new state and history information, as well as newly received response data, navigation commands, etc.

[0325] Returning to FIG. 29, during the interaction (and after completion of the interaction, if necessary based on processing requirements), client response data is collected (at 2940). This data includes video / visual information and speech / audio information captured by the client device(s) camera(s) and microphone(s), respectively. Although not discussed in depth, the collected data may also include biometric authentication results, such as via a haptic interface. This collected response data is then used to classify the health condition (at 2950).

[0326] Figure 33 provides a more detailed example process for classification. The model is first retrieved (at 3310) from the model repository. The user data, social data, clinical data, speech and visual data are all provided (at 3330) to the runtime model server(s) for processing. The inclusion of clinical and / or social data sets the current screening or monitoring methodology apart from previous screening or monitoring methodologies.

[0327] This data is pre-processed to remove artifacts, noise, etc. The pre-processed data is also multiplexed (at 3330). The pre-processed and multiplexed data is provided to the model for analysis and to a third-party ASR system (at 3340). The ASR outputs are integrated (if multiple ASR systems are used in concert) and the resulting machine-readable speech data is also provided to the model. The data is then processed by the NLP model (at 3350a), the acoustic model (at 3350b), the video model (at 3350c), and the descriptive features (at 3350d). Each model operates in parallel, where the results from any given model are sent to the other models to coordinate their operation. It is determined (at 3360) whether the modeling is complete. Due to the fact that model results are interdependent on other model results, the process of modeling is cyclical, whereby in some cases models are adjusted with the results of other models (at 3370) and the modeling process repeats until a final result is determined.

[0328] FIG. 34 describes the process of model tuning in more detail. Model tuning basically includes three sub-processes that run in parallel or otherwise interleaved. These include construction of an NLP model using the results of the acoustic and video models in addition to the descriptive features (at 3371), construction of an acoustic model using the results of the NLP and video models in addition to the descriptive features (at 3372), and construction of a video model using the results of the acoustic and NLP models in addition to the descriptive features (at 3373). As mentioned before, this tuning is not a clearly ordered process. This is because, for example, intermediate results from the acoustic model can be used to tune the NLP model, whose output may affect the video model, which in turn tunes the acoustic model, which in turn tunes the NLP model based on the updated acoustic model results. This may result in a looped computation process, where in each iteration the results are adjusted to be a little more accurate than the previous iteration. An artificial cutoff is imposed on such computation loops to avoid infinite cycling and breakdown of the system due to resource waste. These cutoffs are based on the number of loop cycles or the degree of change in values ​​between one loop cycle and the next. Over time, the results from one loop cycle to the next become closer and closer to each other. At some point, further loop cycles are undesirable due to diminishing returns to model accuracy in consumed processing resources.

[0329] An example of this type of adjustment is when the NLP model determines that the user is not speaking. This result is used by the video model to process individual facial features based on mouth boundaries and eye boundaries. However, when the user is speaking, the video model uses this result to change the model for emotion recognition to not rely on the user's mouth region, but rather on the eye region of the user's face. This is just one simplified example of one type of model adjustment, as it is not limiting.

[0330] Returning to FIG. 33, after the modeling is completed, each model is combined (fused) (at 3380) by weighting the classification results by the time domain. This sub-process is described in more detail in relation to FIG. 35. As mentioned before, one model may be relied on more than another model due to classification confidence or based on events in the response. The clearest example of this is that if there is a period where the user is not speaking, the NLP model classification over this period should be minimized, whereas the weights in the video modeling and the acoustic modeling should be given much more weight. Similarly, if two models suggest that the third model is incorrect or wrong due to fraud or other dissonance, the classification of the odd model may also be weighted accordingly lower than the other models.

[0331] In FIG. 35, this weighting process involves starting with a base weight for each model (at 3381). Next, the response is divided into discrete time segments (at 3382). The length of these time segments is configurable, and in one embodiment, the length of the time segments is set to a value of 3 seconds since most verbal concepts are formed over this length of time. Next, the base weight of each model is modified (at 3383) for each period based on the model confidence level. For example, if an NLP model classifies the first 6 seconds as being 96% confident, and the next 12 seconds as being only 80% confident, then a high weight is applied to the first two periods and a low weight is applied for the next four periods.

[0332] The system also determines (at 3384) when the user is not speaking, typically by relying on the ASR output. During these periods, the NLP model is not useful in determining the user's classification, so the weight of the NLP model is reduced (at 3385) over these periods. The degree of reduction may vary based on configuration, but in some embodiments, NLP is given no weight over periods when the user is not speaking.

[0333] Similarly, based on the features and conclusions from the video and acoustic models, periods may be identified where the patient exhibits voice-based biomarkers associated with dishonesty (at 3386). Excessive fidgeting, gaze shifts, higher pitch, and mumbling may all be correlated with dishonesty, and if multiple features are present simultaneously, the system flags these periods of interaction as suspicious. During such periods, the weights of the NLP model are again decreased (at 3387), but only slightly. Even if the user is not completely honest, the words the user speaks still contain useful information, especially for diagnosing depression. For example, even if the user is dishonest about suicidal thoughts (determined by semantic analysis), syntactic features may still be useful in determining the user's classification. Thus, while weights are reduced during periods of dishonesty, the decrease is typically a quarter of the weight decrease, as opposed to a more abrupt weight decrease.

[0334] After all weight adjustments have been made, the system performs a weighted average of the classification results of the models across the entire response period (at 3388). The end result of this condensation of classifications over time across the various component models results in the fusion model output.

[0335] Returning to FIG. 33, this fusion model output results in a final classification for the interaction (at 3390). The classification, model results, and features are output (at 3399), either in aggregate or in part. Returning to FIG. 29, these results are then presented (at 2960) to clients and other interested parties. This includes selecting which results any given entity should receive. For example, a client may be provided with only the classification results, while a physician for the client receives features related to mood, topics of interest, signs of self-harm or suicidal ideation, etc. In response, an insurance company may receive the classification results as well as a sampling of clinical data related to an individual's risk factors.

[0336] Even after reporting the classification results, the process continues by collecting new information as it becomes available and retraining the model to ensure the highest level of accuracy, as well as by analyzing subsequent interactions and interaction results.

[0337] 36, one example implementation of the acoustic modeling process 3350b is presented in more detail. It should be noted that, despite the enhanced detail in this example of the process, this is a significant simplification of just one method of analysis and is intended only as an example process for clarity and not as a limitation on the analysis performed on the response data.

[0338] In this example process, a variable cutoff value is determined (at 3605) from a training data set. In this particular analysis, the received acoustic signal is converted (at 3610) into a spectrogram image, which provides information about the frequencies of the audio signal and the amplitude at each of these frequencies, and which tracks these over time. In this example process, a sample of the spectrogram image corresponding to a set time is taken (at 3615). In some cases, this may be a 10 second sample of spectrogram data.

[0339] The image is converted into a matrix. This matrix is ​​used in an equation to represent higher-level features. The equation is created from training data using machine learning techniques. The equation includes unknown variables in addition to an input matrix of higher-level features (here, spectrogram image samples). These unknown variables are multiplied, divided, added, or subtracted from the feature matrix (or any combination thereof). Since the solution to the equation is also known, it is necessary to randomly select values ​​for the unknown variables (at 3620) in an attempt to solve the equation (at 3630) to obtain a solution similar to the known solution.

[0340] The difference between the values ​​of the solved equations is compared (at 3630) to the known solution values ​​to calculate an error. This process is repeated thousands or even millions of times until approximations of the correct variable values ​​are found, as determined by a sufficiently low error calculation (at 3635). Once these sufficiently accurate values ​​are found, they are compared (at 3640) to cutoff values ​​originally determined from the training data. If the value is above or below the cutoff, this indicates the presence or absence of a classification, based on the equation utilized. In this manner, a classification of the spectrogram analysis can be determined (at 3645), which can then be output (at 3650) for collaboration with other model results.

[0341] The modeling system logic 5320 includes speech recognition 2210 (FIG. 22), which is shown in more detail in (FIG. 37). The speech recognition is specific to a particular language of speech. Thus, the speech recognition 2210 includes a language-specific voice recognition 3702, which in turn includes several language-specific speech recognition engines 3706A-Z. The particular languages ​​of the language-specific speech recognition engines 3706A-Z shown in (FIG. 14) 7) are merely illustrative examples.

[0342] The speech recognition 2210 also includes a translation engine 3704. For example, assume that the patient speaks a language that is recognized by any of the language-specific speech recognition engines 3706A-Z but not processed by the language model 2214 (FIG. 22). The language-specific speech recognition 3702 (FIG. 37) generates text in the language that the patient speaks, i.e., the patient's language, from the speech signal received from the patient. To enable application of the language model 2214, which in this illustrative example cannot process text in the patient's language, the translation engine 3704 translates the text from the patient's language into a language that can be processed by the language model 2214, e.g., English. Although the language model 2214 may not be as accurate when relying on translation by the translation engine 3704, the accuracy of the language model 2214 is very good with currently available translation technology. Furthermore, the importance of the language model 2214 is significantly diluted by the incorporation of the acoustic model 2218, the visual model 2222, and the clinical data 2220 in the creation of the composite model 2204. As a result, the composite model 2204 is highly accurate despite its reliance on the translation engine 3704 .

[0343] The modeling system logic 5320 includes language model training 2212 (FIG. 22) and language models 2214, which are shown in detail in FIG. 10 and FIG. 11, respectively. The language model training 2212 (FIG. 38) includes logic for training each model of the language model 2214. For example, the language model training 2212 (FIG. 38) includes syntactic language model training 3802, semantic pattern model training 3804, speech fluency model training 3806, and non-verbal model training 3808, which include logic for training syntactic language model 3902, semantic pattern model 3904, speech fluency model 3906, and non-verbal model 3908, respectively, of the language model 2214.

[0344] Each model 3902-3908 contains deep learning (also known as deep structured learning or hierarchical learning) logic to assess a patient's depression from text received from speech recognition 2210.

[0345] The syntactic language model 3902 assesses the patient's depression from syntactic features of the patient's speech. Examples of such syntactic features include sentence length, sentence completion, sentence complexity, and negation. If the patient speaks in short sentences, does not complete sentences, speaks in simple sentences, and / or uses relatively frequent negations (e.g., "no," "not," "couldn't," "won't," etc.), the syntactic language model 3902 determines that the patient is likely to be depressed.

[0346] The semantic pattern model 3904 assesses the patient's depression from the positive and / or negative content of the patient's speech, i.e., from the emotions expressed by the patient. Some studies suggest that the expression of negative thoughts may indicate depression, and the expression of positive thoughts may counter-indicate depression. For example, "the commute here is awful" may be interpreted as an indicator of depression, whereas "the commute here is great" may be interpreted as a counter-indicator of depression.

[0347] The speech fluency model 3906 assesses the patient's depression from the patient's speech fluency characteristics, i.e., speech flow. Fluency characteristics may include, for example, word rate, frequency and duration of speech pauses, prevalence of filler expressions such as "um" and "hmm", and packet speech patterns. Some studies suggest that slower word rates, frequent and / or long speech pauses, and a high incidence of filler expressions may indicate depression. Perhaps more so than other language models 2214, the speech fluency model 3906 may be specific to individual patients. For example, speech rates (word rates) vary widely across geographic regions. The usual speech rate of a patient in New York City may be significantly greater than the usual speech rate of a patient in Minnesota.

[0348] The non-verbal model 3908 assesses the patient's depression from non-verbal characteristics of the patient's speech, such as laughter, chuckles, and sighs. Some studies suggest that sighs may indicate depression, whereas laughter and chuckles (and other forms of partially suppressed laughter, such as giggles) may indicate depression.

[0349] The modeling system logic 5320 includes acoustic model training 2216 (FIG. 22) and acoustic model 2214, which are shown in more detail in FIG. 12 and FIG. 13. Acoustic model training 2216 (FIG. 40) includes logic for training each model of acoustic model 2218 (FIG. 41). For example, acoustic model training 2216 (FIG. 40) includes pitch / energy model training 4002, quality / voicing model training 4004, speech flow model training 4006, and articulatory coordination model training 4008, which include logic for training pitch / energy model 4102, quality / voicing pattern model 4104, speech flow model 4106, and articulatory coordination model 1308, respectively, of acoustic model 2218.

[0350] Each of the models 4102-4108 includes deep learning (also known as deep structured learning or hierarchical learning) logic to assess a patient's depression from the collected patient data 2206 (FIG. 22) and audio signals representing the patient's speech received from preprocessing 2208.

[0351] The pitch / energy model 4102 assesses the patient's depression from the pitch and energy of the patient's speech. Examples of energy include, for example, volume and syllable rate. When the patient speaks with a lower pitch, softer, and / or slower, the pitch / energy model 4102 determines that the patient is likely to be in a depressed state.

[0352] The quality / speech model 4104 assesses a patient's depression from the quality and vocal aspects of the patient's speech. Different sound source changes may occur in depression, affecting voice-related aspects of speech, both in general and specific speech sounds.

[0353] The speech flow model 4106 assesses the patient's depression from the patient's speech flow. Speech flow characteristics can include word rate, frequency and duration of speech pauses, prevalence of filler expressions such as "um" and "hmm", and packet speech patterns.

[0354] The articulatory coordination model 4108 assesses the patient's depression from the articulatory coordination in the patient's speech. Articulatory coordination refers to the fine-tuning of timing between articulatory and source characteristics. This coordination is worsened when the patient is in a depressed state.

[0355] Modeling system logic 5320 (FIG. 53) includes visual model training 2224 (FIG. 22) and visual model 2226, which are shown in more detail in FIG. 14 and FIG. 15, respectively. Visual model training 2224 (FIG. 42) includes logic for training each model of visual model 226 (FIG. 53). For example, visual model training 2224 (FIG. 42) includes facial cues model training 4202 and eye / gaze model training 4204, which include logic for training facial cues model 4302 and eye / gaze model 4304, respectively, of visual model 2226.

[0356] Each of the models 4302-4304 includes deep learning (also known as deep structured learning or hierarchical learning) logic to assess a patient's depression from the collected patient data 2206 (FIG. 22) and video signals representing the patient's speech received from preprocessing 2208.

[0357] Facial cue model 4302 assesses a patient's depression from facial cues recognized in videos of the patient's speech. Eye / gaze model 4304 assesses a patient's depression from observed and recognized eye movements in videos of the patient's speech.

[0358] As previously discussed, the composite model builder 2222 (FIG. 22) builds the composite model 2204 by combining the language model 2214, the acoustic model 2218, and the visual model 2226 to train the combined model using both the clinical data 2220 and the collected patient data 2206. As a result, the composite model 2204 assesses a patient's depression using a combination of what the patient says, how the patient says it, and contemporaneous facial and eye expressions. As such, it provides a particularly accurate and effective tool for assessing a patient's depression.

[0359] The above description is illustrative and not limiting. For example, while certain mental health conditions are addressed by the systems and methods described herein, it should be understood that the techniques described herein can effectively assess and / or screen for several other mental health conditions, such as anxiety, post-traumatic stress disorder (PTSD), and stress in general, drug and alcohol addiction, bipolar disorder, among others. Furthermore, while the assessment test administrator 2202 is described as assessing the mental health of a human subject, who may be a patient, it is understood that "assessment" may refer to a specialized assessment performed by a specialized clinician. As used herein, the assessment performed by the assessment test administrator 2202 may be any type of assessment in the general sense, including screening or monitoring. Scoring The models described herein may generate scores at various stages of the assessment. The scores generated may be scaled scores or binary scores. Scaled scores may span multiple values, while binary scores may be one of two discrete values. The disclosed system may exchange binary scores and scaled scores at various stages of the assessment to monitor different mental states, or update a particular binary score and a particular scaled score for a particular mental state during the course of the assessment.

[0360] The scores generated by the system are either binary or scaled and may be generated after each response to each query in the assessment, or may be formed in part based on previous queries. In the latter case, each marginal score acts to fine-tune the prediction of depression or another mental state and make the prediction more robust. The marginal prediction may thus increase the confidence index of the mental state prediction after a certain number of queries and responses (correlated with a particular intermediate mental state).

[0361] In the case of scaled scores, refining the score allows clinicians to more accurately determine the severity of one or more mental conditions a patient is experiencing. For example, when observing multiple intermediate depression states, refining the scaled score allows clinicians to determine whether a patient is mildly, moderately, or severely depressed. Performing multiple scoring iterations may also help clinicians and administrators to remove false negatives by adding redundancy and robustness. For example, initial mental state predictions may be noisy because relatively few speech segments are available for analysis, and NLP algorithms may not have enough information to determine the semantic context in the patient's recorded speech. Even if a single marginal prediction is itself a noisy estimate, refining the prediction by adding more measurements may reduce the overall variance of the system, thereby resulting in a more accurate prediction. The predictions described herein may be more practical than predictions that can be obtained by simply conducting a survey, since people may have an incentive to lie about their condition. Conducting a survey may result in a large number of false positive and false negative results, which may allow patients in need of treatment to slip through the cracks. Furthermore, trained clinicians may be aware of voice and face-based biomarkers, but may not be able to analyze the large amounts of data that the disclosed system can analyze.

[0362] Scaled scores may be used to represent the severity of a mental condition. Scaled scores may be, for example, a number between 1 and 5 or 0 and 100, where a higher number indicates a more severe or acute form of the mental condition experienced by the patient. Scaled scores may include whole numbers, percentages, or decimals. Conditions for which scaled scores may represent severity include, but are not limited to, depression, anxiety, stress, PTSD, phobias, and panic disorders. In one example, a score of 0 on a depression-related aspect of the assessment may indicate no depression, a score of 50 may indicate moderate depression, and a score of 100 may indicate severe depression. Scaled scores may be a combination of multiple scores. A mental condition may be represented as a composition of mental sub-states, and the patient's composite mental condition may be a weighted average of the individual scores from the mental sub-states. For example, a component score for depression may be a weighted average of the individual scores for anger, sadness, self-image, self-esteem, stress, loneliness, isolation, and anxiety.

[0363] The scaled score may be generated using a model that uses a multi-label classifier. The classifier may be, for example, a decision tree classifier, a nearest neighbor classifier, or a neural network-based classifier. The classifier may generate multiple labels for a particular patient at intermediate or final stages of the assessment, where the labels indicate the severity or degree of a particular mental condition. For example, the multi-label classifier may output multiple numerical values, which may be normalized into probabilities using a softmax layer. The label with the highest probability may indicate the severity of the mental condition experienced by the patient.

[0364] The scaled score may be determined using a regression model. The regression model may determine a fit from training examples expressed as a sum of weighted variables. The fit may be used to estimate a score from a patient with known weights. The weights may be based in part on features that may be derived in part from audiovisual signals (e.g., voice-based biomarkers) and in part from patient information such as patient demographics. The weights used to predict the final score or intermediate scores may be obtained from previous intermediate scores.

[0365] The scaled score may be scaled based on the confidence level. The confidence level may be determined based on the recording quality, the type of model (e.g., audio, visual, semantic) used to analyze the patient's speech from the recording, a time analysis related to the model most frequently used during a particular time period, and the time point of a particular voice-based biomarker within the audiovisual sample. To determine an intermediate score, multiple confidence indices may be obtained. The confidence indices during evaluation may be averaged to determine the weighting of a particular scaled score.

[0366] The binary score may reflect a binary outcome from the system. For example, the system may classify the user as either depressed or not depressed. The system may do this using a classification algorithm such as a neural network or an ensemble method. The binary classifier may output a numeric value between 0 and 1. If the patient's score is above a threshold (such as 0.5), the patient may be classified as "depressed." If the patient's score is below the threshold, the patient may be classified as "not depressed." The system may generate multiple binary scores for multiple intermediate states of the assessment. The system may weight and sum the binary scores from the intermediate states of the assessment to generate an overall binary score for the assessment.

[0367] The output of the models described herein may be converted into a calibrated score, e.g., a score with a unit range. The output of the models described herein may additionally or alternatively be converted into a score with a clinical value. The score with a clinical value may be a qualitative diagnosis (e.g., high risk of severe depression). Alternatively, the score with a clinical value may be a normalized qualitative score that is normalized for the general population or a specific subpopulation of patients. The normalized qualitative score may indicate a risk percentage for the general population or subpopulation.

[0368] The systems described herein may be capable of identifying a mental condition (e.g., a mental or behavioral disorder) in a subject with less error (e.g., 10% less) or with greater accuracy (e.g., 10% more) than a standardized mental health questionnaire or testing tool. The error rate or accuracy may be established against a benchmark standard that can be used by an entity to identify or assess one or more medical conditions that constitute the mental condition. The entity may be a clinician, a medical professional, an insurance company, or a government regulatory agency. The benchmark standard may be an independently verified clinical diagnosis. Reliability The models described herein may use a confidence index. The confidence index may be an indicator of how effective the score generated by the machine learning algorithm is to accurately predict a mental condition such as depression. The confidence index may vary depending on the conditions under which the score was obtained. The confidence index may be expressed as an integer, a decimal, or a percentage. The conditions may include the type of recording device, the ambient space in which the signal was obtained, the background noise, the specificity of the patient's speech, the fluency of the speaker's language, the length of the patient's response, the assessed truthfulness of the patient's response, and the frequency of incomprehensible words and phrases. In situations where the quality of the signal or speech makes the speech more difficult to analyze, the confidence index may be smaller. In some embodiments, the confidence index may be added to the score calculation by weighting the calculated binary score or scaled score with the confidence index. In other embodiments, the confidence index may be provided separately. For example, the system may tell the clinician that the patient has a depression score of 0.93 with 75% confidence.

[0369] The confidence level may be based on the quality of the indicators in the training data used to train the model to analyze the patient's speech. For example, if the indicators are based on a survey or questionnaire completed by the patient rather than an official clinical diagnosis, the quality of the indicator may be determined to be low and therefore the confidence level of the score may be low. In some cases, it may be determined that there is some degree of dishonesty in the survey or questionnaire. In such cases, the quality of the indicator may be determined to be low and therefore the confidence level of the score may be low.

[0370] Various measures may be taken by the system to improve the confidence index, especially if it is affected by the environment in which the evaluation is performed. For example, the system may use one or more signal processing algorithms to filter out background noise, or may use impulse response measurements to determine how to remove the effects of reverberation caused by objects and features in the environment in which the speech sample was recorded. The system may use semantic analysis to find context clues and determine the identities of missing or unintelligible words.

[0371] Additionally, the system may use user profiles to group people based on attitudes, ethnic background, gender, age, or other categories, and because similar groups of people may have similar voice-based biomarkers, the system may be able to predict depression with greater confidence because people who exhibit similar voice-based biomarkers may exhibit depression in a similar manner.

[0372] For example, depressed people from various backgrounds may be classified differently by slow speech, monotonous or low pitch variations, excessive pauses, vocal timbre (gravelly or hoarse), incoherent speech, disorganization or loss of focus, brief responses, and stream-of-consciousness discourse. These voice-based biomarkers may belong to one or more segments of analyzed patients.

[0373] The screening system data store 410 (shown in detail in FIG. 44 ) stores and maintains all user and patient data required for and collected by screening or monitoring in the manner described herein. The screening system data store 410 includes a data store logic 4402, a label estimation logic 4404, and a user-patient database 4406. The data store logic 4402 controls access to the user-patient database 4406. For example, the data store logic 4402 stores audiovisual signals of patient responses and provides patient medical history data upon request. If the requested patient medical history data is not available in the user-patient database 4406, the data store logic 4402 retrieves the patient medical history data from the clinical data server 106. If the requested patient social data is not available in the user-patient database 4406, the data store logic 4402 retrieves the patient social history data from the social data server 108. Non-patient users include healthcare service providers and payers.

[0374] The social media server 108 may contain a wide variety of patient / subject data, including but not limited to retail purchase records, legal records (including criminal records), and income history, as these can provide valuable insight into a person's health. In many cases, these social determinants of disease contribute more to a person's morbidity than medical care. Appendix B provides an "Overview of Health Policy: Relative Contribution of Multiple Determinants to Health Outcomes."

[0375] The sign inference logic 4404 includes logic for specifying the signs for which the various learning machines of the health screening or monitoring server 102 will screen. The sign inference logic 4404 includes a user interface that allows a human operator of the health screening or monitoring server 102 to configure and adjust such signs.

[0376] The label estimation logic 4404 also controls the quality of model training by, among other things, determining whether the data stored in the user-patient database 4406 is of suitable quality for model training. The label estimation logic 4404 includes logic for automatically identifying or modifying labels. In particular, if model training reveals significant data points not yet identified as labels, the label estimation logic 4404 looks for correlations between the data points and patient records, system predictions, and clinical insights to automatically assign labels to the data points.

[0377] Although the interactive screening or monitoring server logic 502 is described as engaging in an interactive verbal conversation with the patient to assess the patient's health status, the interactive screening or monitoring server logic 502 may also operate in a passive listening mode in which the interactive screening or monitoring server logic 502 passively listens to the patient speak without prescribing questions to be asked of the patient.

[0378] In this exemplary embodiment, the passive listening mode has two (2) variants. In the first, "conversational" variant, the patient is involved in a conversation with another person where no part of the conversation is controlled by the interactive screening or monitoring server logic 502. Examples of conversational passive listening include a patient talking to a clinician and a patient talking on the phone reminding the patient of an appointment with the clinician or such a patient talking with a pharmacist about medication. In the second, "fly-on-the-wall" (FOTW) or "ambient" variant, the patient is talking alone or in a public or semi-public place. Examples of ambient passive listening include a person talking in a public place or in a hospital emergency room, or a person talking alone, for example in an audio diary or leaving a phone message. One potentially useful scenario for screening or monitoring a person talking alone involves the interactive screening or monitoring server logic 502 screening or monitoring calls to police emergency services (i.e., "9-1-1"). Analysis of emergency service callers can accurately distinguish emergency callers from non-emergency callers.

[0379] It should be noted that this detailed description is intended to describe what is technically possible. Practice of the technology described herein should comply with legal requirements and restrictions, including federal, state, and / or local laws, which may vary from jurisdiction to jurisdiction. For example, some jurisdictions may require explicit notice and / or consent of the involved party(s) prior to capturing speech. Additionally, acquisition, storage, and retrieval of clinical records should be performed in a manner that complies with applicable jurisdictional requirements(s).

[0380] Patient screening or monitoring system 100B (FIG. 45) illustrates a passive listening variation of patient screening or monitoring system 100 (FIG. 1). Patient screening or monitoring system 100B (FIG. 45) includes a health screening or monitoring server 102, a clinical data server 106, and a social data server 108, which are connected to each other via a WAN 110 as previously described and illustrated.

[0381] Because the patient and clinician are in physical proximity to one another for conversational passive listening, the remaining components of the patient screening or monitoring system 110B are connected to one another and to the WAN 110 via a local area network (LAN) 4510.

[0382] There are several ways to distinguish the patient voice from the clinician voice.

[0383] A particularly convenient method is to have two (2) separate listening devices 4512, 4514, one for the patient and one for the clinician. In this exemplary embodiment, the listening devices 4512, 4514 are smart speakers, such as the HomePod™ smart speaker available from Apple Computer, Cupertino, Calif., the Google Home™ smart speaker available from Google LLC, Mountain View, Calif., and the Amazon Echo™ available from Amazon, Inc., Seattle, Wash. In other embodiments, the listening devices 4512, 4514 may be other types of listening devices, such as, for example, a microphone coupled to the clinician device 114B.

[0384] In some embodiments, a single listening device 4514 is used, and the screening or monitoring server 102 uses conventional voice recognition technology to distinguish between the patient and the clinician. The accuracy of such voice recognition may be improved by training the screening or monitoring server 102 to recognize the clinician's voice prior to any session with the patient. The following description refers to the clinician speaking to the patient, but it is understood that the clinician may be substituted for another clinician. For example, in a call placed by an administrator of the medical institution, such as the clinician's support staff, to a patient, the administrator plays the role of the clinician as described in the context of conversational passive listening. Similarly, in a call placed by a pharmacy to a patient regarding a prescription, a person making a call on behalf of the pharmacy or an automated machine caller plays the role of the clinician as described herein. Appendix C illustrates an exemplary question bank in some embodiments of the present invention.

[0385] Processing by the interactive health screening or monitoring logic 402, and in particular the generic dialogue flow logic 602 (FIG. 7), in conversational passive listening is illustrated by logic flow diagram 4600 (FIG. 46). FIG. 46 illustrates an instantiation of a dynamic mode, where query content is analyzed in real time. Loop step 4602 and next step step 4616 define a loop in which the generic dialogue flow logic 602 processes audiovisual signals of a conversation between a patient and a clinician according to steps 4604-4614. Although steps 4604-4614 are illustrated as separate, consecutive steps, these steps are continuously performed by the generic dialogue flow logic 602 concurrently with one another. The loop of steps 4602-4616 is initiated and terminated by the clinician using conventional user interface techniques, for example, using the clinician device 114B (FIG. 45) or the listening device 4514.

[0386] In step 4604 (FIG. 46), the generic dialogue flow logic 602 recognizes a patient question posed by the clinician and sends the question to the runtime model server logic 504 for processing and analysis. The generic dialogue flow logic 602 receives results 1820 related to the audiovisual signals of the clinician's speech, and the results 1820 (FIG. 18) include a textual representation of the clinician's utterance from the ASR logic 1804 along with further information from the descriptive model analysis 1812. This further information includes an identification of various parts of speech of words in the clinician's utterance.

[0387] In step 4606 (FIG. 46), the generic dialogue flow logic 602 identifies the most similar question in the question and dialogue action bank 710 (FIG. 7). If the question recognized in step 4604 is not identical to any question stored in the question and dialogue action bank 710, the generic dialogue flow logic 602 may identify the closest question in the manner described above with respect to question equivalence logic 1104 (FIG. 11) or may identify the question in the question and dialogue action bank 710 (FIG. 7) that is most linguistically similar.

[0388] In step 4608 (FIG. 46), the general dialogue flow logic 602 retrieves the quality of the closest question from the question and dialogue action bank 710, ie, quality 908 (FIG. 9).

[0389] In step 4610 (FIG. 46), the general dialogue flow logic 602 recognizes an audiovisual signal representing the patient's response to the question recognized in step 4604.

[0390] The patient's response is recognized as the patient's speech immediately following the recognized question. The speech may be recognized as the patient's by (i) determining that the voice is picked up more loudly by listening device 4512 than by listening device 4514, or (ii) determining that the voice is different from a voice previously established and recognized as the clinician.

[0391] In step 4612, the generic dialogue flow logic 602 sends the patient's response, along with the context of the clinician's corresponding question, to the runtime model server logic 504 for analysis and evaluation. The context of the clinician's question is important, especially when the semantics of the patient's response is unclear alone. For example, consider a patient answer of simply "yes." That response will be analyzed and evaluated very differently in response to the question, "Were you able to find parking?" than in response to the question, "Have you ever thought about harming yourself?"

[0392] In step 4614, the generic dialogue flow logic 602 reports to the clinician any intermediate analysis received from the results 1820. If the clinician is using the clinician device 114B during the conversation, for example to review the patient's electronic health record, the report may be in the form of animated gauges showing intermediate scores related to some health condition. Examples of animated gauges include steam gauges, i.e., round dial gauges with moving needles, and dynamic histograms such as those found on audio equalizers in sound systems.

[0393] Once the clinician has finished interactive passive listening, processing is completed according to the loop of steps 4602-4616. In step 4618, the interactive screening or monitoring server logic 502 sends a final analysis of the conversation to the clinician. In general, in the context of step 4618, the "clinician" is always a medical professional or the patient's health record.

[0394] Thus, the health screening or monitoring server 102 may passively screen a patient for any of a number of health conditions during a conversation in which the patient may be involved, without requiring a separate explicit screening or monitoring interview of the patient.

[0395] In ambient passive listening, the health screening or monitoring server 102 listens for and processes ambient speech according to logic flow diagram 4700 (FIG. 47). Processing by the interactive health screening or monitoring logic 402, and in particular the generic dialogue flow logic 602 (FIG. 7), in ambient passive listening is illustrated by logic flow diagram 4700 (FIG. 47). Loop step 4702 and next step step 4714 define a loop in which the generic dialogue flow logic 602 processes audiovisual signals of ambient speech according to steps 4704-4712. Although steps 4704-4714 are illustrated as separate successive steps, these steps are continuously performed by the generic dialogue flow logic 602 concurrently with one another. The loop of steps 4702-4714 is initiated and terminated by a human operator of the involved listening device(s), e.g., listening device 4514.

[0396] In step 4704 (FIG. 47), the generic dialogue flow logic 602 captures ambient voices. In test step 4708, the interactive screening or monitoring server logic 502 determines whether the voice captured in step 4704 is spoken by a voice to be analyzed. In ambient passive listening in at least partially controlled areas, many people likely to speak in such areas may be registered with the health screening or monitoring server 102 so that their voices can be recognized. In a school, students may have registered their voices with the health screening or monitoring server 102 upon enrollment.

[0397] In some embodiments, the people whose voices are to be analyzed are admitted students who are recognized by the generic dialogue flow logic 602. In a hospital, hospital personnel may have registered their voices with the health screening or monitoring server 102 upon employment. Additionally, patients within a hospital may register their voices upon first contact, for example, at the information desk or by hospital personnel in the emergency room. In some embodiments, hospital personnel are excluded from the analysis when they are recognized as speakers by the generic dialogue flow logic 602.

[0398] In an emergency room environment, where analysis of voices unknown to the generic dialogue flow logic 602 is important, the generic dialogue flow logic 602 may still track conversations by unknown speakers. Multiple utterances may be recognized by the generic dialogue flow logic 602 as coming from the same individual. The health screening or monitoring server 102 may also determine the approximate location of unknown speakers in an environment with multiple listening devices, for example, by triangulation using the different relative amplitudes and / or relative timing of arrival of captured speech at the multiple listening devices.

[0399] In other embodiments of ambient passive listening where only one person is speaking, the speaker may be asked to identify themselves. Alternatively, in some embodiments, the identity of the speaker may be inferred or may not be of particular importance. In audio diaries, the speaker may be authenticated by the device or may be assumed to be used by the device's owner. In police emergency call triage, the identity of the caller is less important than the quality of the speaker's voice, such as the speaker's location, emotion, energy, and the substantive content of the speaker's utterances.

[0400] In those embodiments where only one person is speaking, the generic dialogue flow logic 602 always determines that the speaker should be analyzed.

[0401] If the speaker is not to be analyzed, the generic dialogue flow logic 602 sends the captured surrounding speech to the runtime model server logic 504 for context processing and analysis. The generic dialogue flow logic 602 receives results 1820 regarding the audiovisual signals of the captured speech, and the results 1820 (FIG. 18) include a textual representation of the captured speech from the ASR logic 1804 along with further information from the descriptive model analysis 1812. This further information includes an identification of the various parts of speech of the words in the clinician's speech. The generic dialogue flow logic 602 processes the captured speech results 1820 to build a context.

[0402] After step 4708 (FIG. 47), the process moves to the loop step 4702 via the next step step 4714, and passive listening continues according to the loop of steps 4702-4714.

[0403] If, in test step 4706, the interactive screening or monitoring server logic 502 determines that the utterance captured in step 4704 was spoken by a voice to be analyzed, then processing transitions to step 4710. In step 4710, the generic dialogue flow logic 602 transmits the captured utterance, along with any runtime determined context that occurs prior to but contemporaneous with step 4708 or step 4710, to the runtime model server logic 504 for analysis and evaluation.

[0404] In step 4712, the generic dialogue flow logic 602 processes any alerts triggered by the analysis resulting from the runtime model server logic 504 according to predefined alert rules. These predefined alert rules are similar to the workflow 4810 described below. Essentially, these predefined alert rules are in the form of if-then-else logic elements that specify logic conditions and corresponding actions to be taken in such conditions.

[0405] The following are examples of alert rules that may be implemented by the interactive screening or monitoring server logic 502: In a police emergency system call where a caller first speaks to an automated triage system and the speech is determined to be highly emotional and anxious, semantically indicative of a highly urgent situation, e.g., a motor vehicle accident with serious injuries, a very high priority may be assigned to the call and it may be prioritized over less urgent callers. In a school hallway where the interactive screening or monitoring server logic 502 recognizes frantic speech and screaming, along with semantic content indicative of the presence of weapons and / or explicit violent acts, the interactive screening or monitoring server logic 502 may trigger immediate notification of law enforcement and school personnel. In a voice diary where the patient is detected to be at least moderately depressed, the interactive screening or monitoring server logic 502 may record the analysis in the patient's clinical record so that the diary entry can be discussed with the patient's behavioral health care provider the next time they meet with the patient. In situations where the triggering condition of the captured speech is particularly severe and urgent, the interactive screening or monitoring server logic 502 may report the speaker's location if it can be determined.

[0406] Processing continues according to the loop of steps 4702-4714 (FIG. 47) until stopped by the interactive screening or monitoring server logic 502 or a human operator of an associated listening device.

[0407] Thus, the health screening or monitoring server 102 may passively screen patients for any of a number of health conditions outside of a one-to-one conversation with a medical professional.

[0408] 4, the healthcare management logic 408 provides expert recommendations in response to the health status analysis of the interactive health screening or monitoring logic 402. The healthcare management logic 408 is shown in further detail in FIG.

[0409] The healthcare management logic 408 includes manual workflow management logic 4802, automated workflow generation logic 4804, workflow execution logic 4806, and workflow configuration 4808. The manual workflow management logic 4802 implements a user interface that allows a human administrator to create, modify, and delete workflows 4810 of the workflow configuration 4808 by physical manipulation of one or more user input devices of a computer system used by the administrator. The automated workflow generation logic 4804 performs statistical analysis of patient data stored in the screening or monitoring system data store 410 to identify workflows to achieve predetermined goals. Examples of such goals include minimizing the projected cost of patient care over the next two years and maximizing the reduction in hemoglobin A1C over one year while also minimizing the cost of initial referrals.

[0410] The workflow execution logic 4806 processes the workflows 4810 of the workflow configuration 4808 , thereby evaluating conditions and executing actions of workflow elements 4820 .

[0411] In some embodiments, the workflow execution logic 4806 processes the workflow 4810 in response to receiving any screening or monitoring final results according to logic flow diagram 800 (FIG. 8), using those results in the workflow's processing conditions.

[0412] The workflow configuration 4808 (FIG. 48) includes data representing a number of workflows 4810. Each workflow 4810 includes workflow metadata 4812 and data representing a number of workflow elements 4820.

[0413] The workflow metadata 4812 is metadata for the workflow 4810 and includes data representing the description 4812, creator 4816, and schedule 4818. The description 4812 is information intended to inform any human operator of the nature of the workflow 4810. The creator 4816 identifies the entity that created the workflow 4810, whether a human administrator or the automated workflow generation logic 4804. The schedule 4818 specifies the dates and times and / or conditions under which the workflow execution logic 4806 is intended to process the workflow 4810.

[0414] Workflow elements 4820 collectively define the behavior of workflow execution logic 4806 when processing a workflow. In this exemplary embodiment, the workflow elements are each one of two types: a condition, such as condition 4900 (FIG. 49), and an action, such as action 5000 (FIG. 50).

[0415] In this exemplary embodiment, condition 4900 specifies a Boolean test that includes an operand 4902, an operator 4904, and another operand 4906. In this exemplary embodiment, operator 4904 may be any of several Boolean test operators, such as, for example, =, ≠, >, ≧, <, and ≦. Each of operands 4902, 4906 may be result 1820 (FIG. 18) or any portion thereof, a constant, or null. As a result, any result of a given screening or monitoring, e.g., result 1820, any information about a given patient stored in screening or monitoring system data store 410, and any combination thereof, may be any of operands 4902, 4906.

[0416] The next workflow element(s) 4908 specify one or more workflow elements to process if the test of operands 4902, 4906 and operator 4904 evaluates to a true Boolean value, and the next workflow element(s) 4910 specify one or more workflow elements to process if the test of operands 4902, 4906 and operator 4904 evaluates to a false Boolean value.

[0417] Each of the next workflow element(s) 4908, 4910 may be either a condition, an action, or null. By accepting conditions such as condition 4900 in the next workflow element(s) 1908, 4910, complex tests using AND and OR operations can be expressed in workflow element 4820. In another embodiment, condition 4900 may include more operands and operators combined with AND, OR, and NOT operations.

[0418] Because each of operands 4902, 4906 can be null, condition 4900 may test for the mere presence or absence of occurrences in a patient's data. For example, to determine if a patient has had a hemoglobin A1C blood test, condition 4900 may determine if the most recent hemoglobin A1C test result is null. If equal, the patient has not had a hemoglobin A1C blood test at all.

[0419] Action 5000 (FIG. 50) includes action logic 5002 and one or more of the following workflow elements 5004. Action logic 5002 represents the substantive action to be taken by workflow execution logic 4806, and generally creates or recommends a particular course of action in the care of the patient, which may range from a particular treatment protocol to a more holistic paradigm. Examples include referring the patient to a healthcare provider, enrolling the patient in a particular care program, and recording the recommendation in the patient's file so that the patient's clinician can review the recommendation at their next visit. Examples of referring a patient to a care provider include referring the patient to a psychiatrist, medication management coach, physical therapist, nutritionist, fitness coach, dietician, social worker, etc. Examples of enrolling a patient in a program include a telepsychiatry program, a group therapy program, etc.

[0420] Examples of recommendations recorded in a patient's file include recommended changes to medications, whether that be a change to a particular medication prescribed or simply a change in the dosage of a medication already prescribed to the patient, and other treatments or not. Additionally, referrals and enrollments may be influenced by recommendations regarding referrals and enrollments in the patient's file, thereby allowing clinicians to make the final decisions regarding the patient's care.

[0421] As previously described, the automated workflow generation logic 4804 (FIG. 48) performs statistical analysis of patient data stored in the screening or monitoring system data store 410 to identify workflows to achieve predetermined goals. Examples of such goals given above include minimizing the projected cost of the next two (2) years of care for the patient, while also maximizing the reduction in hemoglobin A1C over one year, while minimizing the cost of initial referrals. The automated workflow generation logic 4804 is described in a first exemplary context, namely, minimizing the projected cost of the next two (2) years of care for the patient.

[0422] The manner in which the automated workflow generation logic 4804 identifies a workflow to achieve a given goal is illustrated by logic flow diagram 5100 (FIG. 51).

[0423] The automated workflow generation logic 4804 includes deep learning machine logic. In step 5102, a human computer engineer configures this deep learning machine logic of the automated workflow generation logic 4804 to analyze patient data from the screening or monitoring system data store 410 in association with indicators specified by a user, such as indicators related to the cost of care for each patient over a two-year period in this illustrative example. Users of the health screening or monitoring server 102, who are not just patients, are typically either medical professionals or health care payers. In either case, information about events in a given patient's health care history is available and included in the automated workflow generation logic 4804 by the human engineer so that the automated workflow generation logic 4804 can track the cost of care for the patient from the patient's medical records.

[0424] Further, in step 5102, the human engineer trains the deep learning machine logic of the automated workflow generation logic 4804 using all relevant data in the screening or monitoring system data store 410. After such training, the deep learning machine logic of the automated workflow generation logic 4804 includes a highly complex decision tree that predicts costs for each patient over a two-year period.

[0425] In step 5104, the automated workflow generation logic 4804 determines which events in the patient's medical history have the most impact on the cost of care for the patient over a two-year period for a statistically significant portion of the patient population. In particular, the automated workflow generation logic 4804 identifies deep learning machine (DLM) nodes of a decision tree that have the most impact on a given goal, e.g., the cost of care for the patient over a two-year period. There are several known techniques for making such decisions automatically, and the automated workflow generation logic 4804 implements one or more of them to identify these important nodes. Examples of techniques for identifying highly influential events / decisions ("nodes" in machine learning terminology) in deep learning machines include, for example, random decision forests (supervised or unsupervised), multinomial logistic regression, and naive Bayes classifiers. These techniques are known and will not be described herein.

[0426] Loop step 5106 and next stage step 5112 define a loop in which the automated workflow generation logic 4804 processes each of the influential nodes identified in step 5104. The particular node processed by the automated workflow generation logic 4804 in a given iteration of the loop of steps 5106-5112 may be referred to as the target node.

[0427] In step 5108, the automated workflow generation logic 4804 forms a condition, e.g., condition 4900 (FIG. 49), from the internal logic of the subject node. The internal logic of the subject node receives data representing one or more events in a patient's history and / or one or more phenotypes of the patient and makes one or more decisions representing branching to other nodes. In step 5108 (FIG. 51), the automated workflow generation logic 4804 generalizes the data received by the subject node and the internal logic of the subject node that maps the received data to a decision.

[0428] In step 5110, the automated workflow generation logic 4804 forms an action, e.g., action 5000 (FIG. 50), according to the branch from the target node that ultimately results in the best outcome associated with the given goal, e.g., lowest cost over a two year period. The conditions formed in step 5108 (FIG. 51) and the actions formed in step 5110 collectively form the workflow generated by the automated workflow generation logic 4804.

[0429] When all influential nodes have been processed according to the loop of steps 5106-5112, processing by the automatic workflow generation logic 4804 is completed, thereby forming several workflows.

[0430] In this exemplary embodiment, the automatically generated workflows are subject to human approval before they are actually deployed within the healthcare management logic 408. In another embodiment, the healthcare management logic 408 automatically deploys the workflows automatically generated by the automated workflow generation logic 4804, but limits the actions to only recommendations to healthcare professionals. It is technically feasible to fully automate the generation of workflows and changes in patient care without any human oversight; however, doing so may be contrary to public health policy as it is implemented today. Clinical scenario

[0431] The disclosed system may also be used to assess mental health status from primary care health interactions. For example, the system may be used to enhance inferences regarding a patient's mental health obtained by a trained individual health care provider. The system may be used to assess mental health status from preliminary screening or monitoring calls (e.g., calls made by a prospective patient to a health care provider organization for the purpose of setting up a medical appointment with a trained mental health professional). For primary screening, the health care professional may ask the patient certain questions in a particular order to ascertain the patient's mental health treatment needs. A recording device may record the prospective patient's responses to one or more of these questions. The prospective patient's consent may be obtained before this is done.

[0432] The system may perform speech or semantic analysis on the speech snippets it collects from prospective patients. For example, the system may determine the relative frequency of words or phrases associated with depression. For example, the system may predict that a user is depressed if the user speaks with phrases indicative of suicidal thoughts, instincts to harm themselves, phrases indicative of poor body image or self-image, and terms associated with negative thoughts such as anxiety, isolation, or loneliness. The system may also pick up non-lexical or non-verbal clues to depression, such as pauses, gasps, sighs, and slurred or muttered speech. These terms and non-lexical clues may be similar to those picked up from training examples, such as patients who have completed surveys (such as the PHQ-9).

[0433] The system may determine information about mental health by probing the user's physical health. For example, the user may feel anxious or sad about their physical characteristics or strength. The questions used to elicit information may relate to vitals such as blood pressure, resting heart rate, family history of disease, blood sugar, body mass index, body fat percentage, injuries, deformities, weight, height, eyesight, eating disorders, cardiovascular endurance, diet, or strength. The patient may provide utterances that indicate disappointment, resentment, sadness, or defensiveness. For example, the patient may make excuses for why they have not received medical treatment, why their diet is not working, why they have not started an exercise program, or may speak negatively about their height, weight, or physical characteristics. Such negative expressions about physical health may be correlated with anxiety.

[0434] Models may be continuously active or passive. A passive learning model may not change the way it learns in response to new information. For example, a passive learner may continually use a particular condition to converge to a prediction, even as new types of feature information are added to the system. However, such models may have limited effectiveness if a large amount of training data is not available. In contrast, an active learning model may employ a human to converge more quickly. Active learning may ask humans targeted questions to do this. For example, a machine learning algorithm may be used on a large number of unlabeled voice samples. The algorithm may easily classify some as indicative of depression, while others are ambiguous. The algorithm may ask the patient if they feel depressed when uttering a particular voice segment. Or, the algorithm may ask a clinician to classify the sample.

[0435] The system may be able to perform quality assurance of healthcare providers using the voice biomarkers. Data from the system may be given to healthcare professionals to assist them in detecting lexical and non-lexical cues corresponding to depression in patients. The healthcare professionals may use changes in pitch, vocal rhythm, and vocal tics to determine how to proceed with care. The system also allows healthcare professionals to evaluate which questions elicit responses from patients that are most predictive of depression. The healthcare professionals may use data from the system to cross-train themselves to search for lexical and non-lexical cues and monitor the delivery of care to determine whether it is effective in screening or monitoring patients. For example, a healthcare professional may be able to observe a second healthcare professional asking questions of a subject to determine whether the second healthcare professional is asking questions that elicit useful information from the patient. The healthcare professional may ask questions in person or remotely, such as from a call center. Medical personnel may use the semantic and audio information generated by the system to create a standardized method of eliciting information from patients that yields the most cues from the patient based on that method.

[0436] The system may be used to provide a dashboard that tabulates the voice-based biomarkers observed in a patient. For example, a healthcare professional may be able to track the frequency of certain biomarkers to keep track of a patient's condition. A healthcare professional may be able to track these frequencies in real time to evaluate how they are performing in their treatment. A healthcare professional may also be able to track these frequencies over time to monitor the performance of patients under treatment or in the progress of recovery. Mental health providers may be able to use this collected data to evaluate each other's performance.

[0437] The dashboard may show real-time biomarker data as the fragments are being analyzed. The dashboard may show line graphs showing trends in the measured biomarkers over time. The dashboard may show predictions made at various points in time, thereby charting the patient's progress with respect to treatment. The dashboard may show the patient's response to treatment by various providers.

[0438] The system may be able to translate one or more of its models across different patient settings. This may be done to account for background audio information in various settings. For example, the system may use one or more signal processing algorithms to normalize audio input across settings. This may be done by performing impulse response measurements of multiple locations to determine transfer functions of signals collected at those locations to normalize audio recordings. The system may also account for training at various locations. For example, a patient may feel more comfortable discussing sensitive issues at home or in a therapist's office than over the phone. Thus, the voice-based biomarkers acquired in these settings may be different. The system may be trained at multiple locations, or the training data may be labeled by location before being processed by the system's machine learning algorithms.

[0439] The model may be transferred from location to location, for example, by using signal processing algorithms. The model may also be transferred by modifying the questions asked to the patient based on the patient's location. For example, it may be determined which particular question or set of questions corresponds to a particular response within the context of a particular location. The questions may then be administered by the medical professional in a manner that provides the same response from the patient.

[0440] The system may be capable of training the audio biomarker model using standard clinical encounters. The system may collect records of clinical encounters for somatic complaints. The complaints may be related to injuries, illnesses, or chronic conditions. With the patient's permission, the system may record conversations the patient has with healthcare professionals during appointments. The somatic complaints may indicate the patient's feelings regarding their health condition. In some cases, somatic complaints may cause significant distress to the patient, affecting the patient's overall disposition and causing depression.

[0441] The data may be encrypted when it is collected or during transmission to one or more servers in the system. The data may be encrypted using a symmetric key encryption scheme, a public key encryption scheme, or a blockchain encryption method. The computations performed by one or more machine learning algorithms may be encrypted using a homomorphic encryption scheme, such as a partially homomorphic encryption scheme or a fully homomorphic encryption scheme.

[0442] To protect privacy, the data may be analyzed locally, or the system may analyze the data in real-time by implementing trained machine learning algorithms to operate on speech sample data recorded at the location where the reservation is being made.

[0443] Alternatively, the data may be stored locally. To protect privacy, features may be extracted before being stored in the cloud for later analysis. The features may be anonymized to protect privacy. For example, patients may be given identifiers or pseudonyms to hide their true identities. To ensure that patient identities are not compromised, the data may undergo differential privacy. Differential privacy may be obtained by adding noise to the dataset. For example, the dataset may include 100 records corresponding to 100 user names and the added noise. If an observer has information about 99 records corresponding to 99 users and knows the remaining user names, the observer cannot match the remaining records with the remaining user names because noise is present in the system.

[0444] In some embodiments, a local model may be embedded in a user device. The local model may be capable of performing limited machine learning or statistical analysis, subject to the computational power and storage constraints of the device. The model may be capable of performing digital signal processing on audio recordings from the patient. The mobile device used may be a smartphone or tablet computer. The mobile device may be capable of downloading algorithms over a network for analysis of the local data. The use of a local device may ensure privacy, since the collected and analyzed data may not travel over a network.

[0445] The voice-based biomarkers may be correlated with laboratory or physiological measurements. The voice-based biomarkers may be correlated with mental health-related measurements. For example, the voice-based biomarkers may be compared to the effectiveness of psychiatric treatments or logs taken by medical professionals such as therapists. The voice-based biomarkers may be compared to responses to survey questions to see if the voice-based analysis is consistent with assessments commonly performed in the field.

[0446] The voice-based biomarkers may be associated with physical health-related measurements. For example, vocal issues such as illness may contribute to the patient producing a vocal sound that needs to be taken into account to provide an actionable prediction. Additionally, a prediction of depression over a timescale in which the patient is recovering from an illness or injury may be compared to the patient's health status over that timescale to see if treatment is improving the patient's depression or depression-related symptoms. The voice-based biomarkers may be compared to data related to brain activity collected over multiple time points to determine the clinical effectiveness of the system.

[0447] Training of the model may be continuous, so that the model is running continuously while voice data is collected. Voice-based biomarkers may be added to the system continuously and used for training during multiple time periods. The model may be updated using data as it is collected.

[0448] The system may use a reinforcement learning mechanism, where survey questions may be dynamically altered to elicit voice-based biomarkers that result in a high confidence prediction of depression. For example, the reinforcement learning mechanism may be able to select questions from a group. Based on a previous question or a series of previous questions, the reinforcement mechanism may select questions that may result in a high confidence prediction of depression.

[0449] The system may be able to determine which questions or series of questions may result in a particular elicitation from the patient. The system may use machine learning to predict the particular elicitation, for example by generating a probability. The system may use a softmax layer to generate probabilities for multiple elicitations. The system may use as features not only the specific questions but also the time at which these questions are asked, the period of the study at which the questions are asked, the date and time at which the questions are asked, and the point in time during the course of treatment at which the questions are asked.

[0450] For example, a particular question asked at a particular time about a sensitive subject may elicit crying from the patient, which may be strongly associated with depression. Upon receiving the context that it is a particular time, the system may recommend posing the question to the patient.

[0451] The system may include a method for dynamically influencing the course of treatment using the voice-based biomarkers. The system may record a user's interview over a period of time and determine from the recorded interview whether the treatment has been effective. For example, if the voice-based biomarkers become less indicative of depression over a long period of time, this may be evidence that the prescribed treatment is working. On the other hand, if the voice-based biomarkers become more indicative of depression over a long period of time, the system may prompt a medical professional to pursue a change in treatment or to pursue a current course of treatment more aggressively.

[0452] The system may proactively recommend changes in treatment. In embodiments where the system is continuously processing and analyzing data, the system may detect a sudden increase in a voice-based biomarker indicative of depression, which may occur over a relatively short time frame during the course of treatment. The system may also be able to proactively recommend changes if the course of treatment has not been effective over a certain period of time (e.g., 6 months, 1 year).

[0453] The system may be able to track the probability of a particular response to a medication. For example, the system may be able to track voice-based biomarkers acquired before, during, and after a course of treatment to analyze changes in scores indicative of depression.

[0454] The system may be able to track the probability of a particular patient's response to a medication by having been trained on similar patients. The system may use this data to predict a patient's response based on the responses of patients from similar demographics. These demographics may include age, sex, weight, height, medical history, or a combination thereof.

[0455] The system may be able to track the likelihood of a patient's adherence to a course of medication or treatment. For example, the system may be able to predict whether a treatment is having an impact on a patient based on analysis of a time series of voice-based biomarkers. At this time, a healthcare professional may ask the patient if they are following the treatment.

[0456] Furthermore, the system may be able to know whether the patient is complying with the treatment by analyzing their biomarkers based on the survey questions. For example, the patient may become defensive, take long pauses, stutter, or act in a way that makes it clear that the patient is lying about adhering to the treatment plan. The patient may also express sadness, shame, or regret about not ...

Claims

1. 1. A computer-implemented method for assessing a mental state of a subject, comprising: (a) receiving, by the computer, speech or text data from the subject; (b) processing the data by the computer using one or more individual, combined, or fusion models, including a natural language processing (NLP) model, an acoustic model, and / or a visual model, the one or more individual, combined, or fusion models output a plurality of outputs corresponding to a plurality of time segments of the data, the plurality of outputs being temporally weighted; (c) generating, by the computer, an assessment of the mental state of the subject and a confidence level of the assessment based on at least the plurality of outputs; The method includes:

2. The data in (a) may include the steps of: (i) sending at least one query in audio or text form to the subject, the at least one query configured to elicit at least one verbal response from the subject; and (ii) receiving data from the subject in response to sending the at least one query, the data including the at least one verbal or textual response. The method of claim 1 , obtained by

3. The psychiatric condition is selected from the group consisting of depression, anxiety, stress, post-traumatic stress disorder (PTSD), schizophrenia, suicidality, bipolar disorder, fatigue, loneliness, and lack of motivation. The method of claim 1.

4. The assessment includes a score indicating a level of severity of the psychiatric condition. The method of claim 1.

5. converting said scores into one or more scores with clinical values; The method of claim 4 further comprising:

6. transmitting the assessment to a medical professional for use in assessing the mental state of the subject, the transmitting being performed just-in-time during the assessment or in real-time after the assessment is completed, the assessment being transmitted within a report indicative of the mental state of the subject. The method of claim 1.

7. The one or more individual models, combined models, or fusion models include a composite model that is an aggregation of two or more different models. The method of claim 1.

8. the confidence level is based at least in part on a quality of an audio signal of the data or at least in part on an indication of confidence of a speech recognition process performed on the audio signal of the data; The method of claim 1.

9. the confidence level is based at least in part on a level of truthfulness associated with the assessment, a length and / or duration of the audio or text, a quality indicator of the data associated with metadata of the data, the subject, or a context of the data, or a model-generated confidence from an acoustic, NLP, or speech recognition model; The method of claim 1.

10. The assessment is continually updated with information obtained from each of the subject's responses to a preceding query. The method of claim 1.

11. transmitting the assessment with clinical value to the subject and / or the subject's contacts, health care professionals, during screening, monitoring or diagnosis or after the screening, monitoring or diagnosis has been completed, for use in assessing the subject's mental health and / or in providing mental health care. The method of claim 10.

12. determining that the confidence level does not meet a predetermined criterion; generating, in real time, at least one query; using said at least one query until said confidence level meets said predetermined criteria; sending the at least one query to the subject; receiving further data from the subject, the further data including at least one response; processing, by the computer, the data and the further data using one or more individual, combined or fused models including the natural language processing (NLP) model, the acoustic model and / or the visual model; generating, by the computer, a confidence level of the assessment and an assessment of the mental state of the subject based on at least the plurality of outputs. Repeating the steps of: The method of claim 1 further comprising:

13. extracting one or more topics of interest of the subject from the data using a topic model. The method of claim 1.

14. relating said one or more topics of interest to a mental health condition; Further comprising: The method of claim 13.

15. using said related one or more topics of interest at least in part as a basis for an analysis that is reported to stakeholders; Further comprising: The method of claim 14.

16. (1) selecting a first model that is optimized for sensitivity in assessing whether the subject has the mental condition; or (2) selecting a second model that is optimized for specificity in assessing whether the subject has the mental condition; and processing the data using the selected first model or the second model to generate the assessment; The method of claim 1 further comprising:

17. using metadata to generate the one or more assessments of the mental condition associated with the subject, the metadata including demographic information and / or a medical history of the subject. The method of claim 1.

Images