A voice-based depression intelligent recognition method, early warning system and storage medium

Abstract

The application discloses a kind of depression intelligent recognition method, early warning system and storage medium based on voice, comprising the following steps: presenting the standardization chart of Rorschach Inkblot Test to subject in turn, collecting the voice information of subject, and the original voice is preprocessed, and Rorschach Inkblot Test voice library is constructed.Extract the acoustic feature and input the deep learning model trained in advance, utilize its time series modeling capability to capture the depression related dynamic mode in voice, and the risk score P of the depression state of subject is obtained by deep learning model calculation.Based on risk score P, output depression risk assessment result.The application combines Rorschach Inkblot Test with deep learning technology, through the voice collection and time series modeling of subject, realizes the intelligent recognition early warning of depression state, provides scientific reference basis for large-scale mental health screening, significantly improves the depression early intelligent recognition early warning work efficiency and objectivity.

Description

Technical Field

[0001] This invention relates to the field of computer-aided psychological assessment and mental health early warning technology, and in particular to a voice-based intelligent depression recognition method, early warning system and storage medium. Background Technology

[0002] Depression is a mood disorder characterized by a significant and persistent low mood, and is characterized by its chronic nature and tendency to recur. For certain occupational groups, due to the high intensity of training, strict discipline, high risk, and relatively closed environment, the continuous stress they face is often significantly higher than that of the general population. Depression, as one of the leading causes of suicide among working professionals, has severely hampered the sustainable development of related industries. Therefore, early warning and effective intervention for depressive states in this population are of great practical significance.

[0003] Currently, depression screening primarily relies on self-report scales, such as the Self-Rating Depression Scale (SDS) and the Patient Health Questionnaire-9 (PHQ-9). These methods are convenient to implement, simple to score, and easy to administer on a large scale. However, their inherent technical limitations are also quite significant:

[0004] (1) Highly subjective and easily influenced by social desirability: The results of self-report scales rely entirely on the subjects' self-reports. In high-risk situations involving personal career development and vital interests, such as pre-employment mental health assessments and regular occupational health checkups, test subjects often engage in social desirability processing, tending to present their best side, leading to measurement bias and reducing the reliability and validity of the scale. This results in some employees with pre-existing depressive tendencies not being detected and warned in a timely manner. Untimely detection and intervention of depression not only seriously affect the individual's physical and mental health and daily life, but may also weaken their occupational performance and social function.

[0005] (2) High face validity, easily interfered with by defense mechanisms: The purpose of the scale items is obvious, and the test takers are likely to guess the test intention and thus answer according to social expectations rather than their own true situation. In view of the inertia of middle option preference caused by the "mediocrity thinking" in the Chinese cultural context, and the self-presentation bias caused by the introverted expression of emotions, self-report scales are difficult to effectively avoid the subjective modification of the test takers.

[0006] (3) Limited information dimensions: Traditional scales mainly measure conscious behavior and subjective feelings, making it difficult to explore the deep psychological traits such as personality characteristics, cognitive processing patterns and emotional regulation ability at the subconscious level in a deep and undisturbed manner. These traits are often key factors in the formation and maintenance of depressive state.

[0007] To overcome the limitations of self-report scales and seek a more objective assessment method that can bypass psychological defenses, projective tests have been introduced as a supplementary approach. Among them, the Rorschach Inkblot Test (RIT) is one of the most authoritative and widely used psychological projective tests internationally. By analyzing an individual's response to a fuzzy inkblot, it can effectively overcome psychological defense mechanisms and objectively present the subject's psychological state and potential personality, showing significant advantages over self-report scales. Studies have shown that the Depression Index (DEPI) in the Rorschach Inkblot Test is effective in predicting depression in clinical patients with depression and university students.

[0008] However, applying the Rorschach inkblot test to identify and warn of depression, especially in practical application, faces significant technical challenges and limitations:

[0009] (1) Existing research on depression identification relies on foreign behavioral corpora: Most studies that classify and identify depression based on behavioral data (such as speech, text, and facial expressions) use mature foreign depression corpora. Research using existing data is easily affected by factors such as the background of the test, the data collection conditions, and the language, making it difficult to replicate the research process and generalize the results. Therefore, it is particularly important to create an independent speech database.

[0010] (2) Heterogeneity of self-collected data methods and its limitations on cross-sectional comparison: In studies on depression identification using self-collected data, the data acquisition methods are quite diverse, making cross-sectional comparison difficult. In speech experiments for depression identification, different speech patterns involve different processes of information collection, thought processing, and language organization by the speaker, which will be reflected in changes in speech features. Some researchers collect speech data through self-designed experiments, such as interviews, reading aloud, descriptions, and speeches, making it difficult to compare the results of different experiments and compare the merits of different models.

[0011] (3) Traditional machine learning paradigm dominates psychological speech classification research: Currently, most psychological research on speech feature classification and recognition adopts traditional machine learning methods. Although traditional machine learning has the advantage of training models without large samples, it may also lose key features, making it difficult to judge the quality of features, resulting in poor model performance.

[0012] In summary, a significant contradiction exists in existing technologies: on the one hand, while traditional self-report scales are convenient for large-scale use, they lack objectivity; on the other hand, research on depression identification based on the Rorschach inkblot test suffers from dependence on foreign corpora, heterogeneity in self-collected data methods, and limitations of traditional statistical analysis paradigms, making it difficult to develop a unified and generalizable identification model. There is an urgent need in this field for a novel solution that combines the deep insights of the Rorschach inkblot test with deep learning technology to assist in achieving efficient and objective intelligent identification and early warning of depressive states. Summary of the Invention

[0013] This invention addresses the shortcomings of existing technologies by proposing a voice-based intelligent depression recognition method, early warning system, and storage medium. It solves the technical problem of how to combine the multi-dimensional psychological projection features contained in the Rorschach inkblot test with deep learning technology, overcomes the limitations of existing depression recognition research that relies on foreign corpora and is dominated by traditional statistical analysis, and achieves efficient, objective, and accurate intelligent recognition and early warning of depressive states.

[0014] A speech-based intelligent method for recognizing depression includes the following steps:

[0015] Step S1: Model training step. Based on the Rorschach inkblot test, an RTSC-DEP speech library is constructed. The deep learning model is trained using depression diagnostic labels as supervision signals to obtain a depression early warning model.

[0016] Step S2: Model application step. Present the Rorschach inkblot test cards to the subject, collect voice recordings and input them into the trained depression recognition model, and output the depression risk assessment results.

[0017] In step S1, the method for constructing RTSC-DEP includes the following steps:

[0018] Step S11: Subject screening and enrollment. Initial screening was conducted using the Self-Rating Depression Scale, combined with structured clinical interviews with psychiatrists, and participants were divided into a depression group and a control group.

[0019] Step S12: Voice data collection. The standardized Rorschach inkblot test cards are presented to the participants, and the participants' descriptive voices are collected and recorded throughout the process.

[0020] Step S13: Speech preprocessing, which involves noise reduction, format standardization, and removal of invalid segments from the collected raw speech;

[0021] Step S14: Data labeling, labeling the voice data based on the structured clinical interview diagnostic results of ICD-11.

[0022] In step S11, the participants are selected by “initial screening of scales + doctor’s diagnosis”. The core assessment tools include the Self-Rating Depression Scale (SDS) and the International Classification of Diseases, Eleventh Revision (ICD-11).

[0023] In step S14, the depression group and the control group are strictly divided based on the results of the structured clinical interviews conducted by ICD-11.

[0024] Step S2 specifically includes the following steps:

[0025] Step S21: Data collection and preprocessing step. Present the Rorschach inkblot test cards to the subjects, record the subjects' descriptive speech, and preprocess the raw speech.

[0026] Step S22: Feature extraction step, extracting a set of acoustic features that can represent emotions and cognitive states from the preprocessed speech;

[0027] Step S23: Model recognition step, after the extracted acoustic features are standardized, they are input into the pre-trained deep learning model to calculate the risk score P of the subject belonging to the depressive state;

[0028] Step S24: Result output step, based on the risk score P, output the depression risk assessment result.

[0029] Specifically, step S21, which involves preprocessing the original speech by noise reduction, segmentation, and format standardization, includes the following steps:

[0030] Step S211: Record using a high-fidelity directional microphone and save as a mono audio file in lossless audio format.

[0031] Step S212: Use a noise reduction algorithm based on spectrum analysis to identify and suppress steady-state low-frequency noise, and adaptively adjust the noise reduction parameters according to the ambient noise level.

[0032] Step S213: Manually edit out the experimenter's intervention segment, closed-ended question and answer segment and other high-frequency noise, and retain the subject's active expression segment and the subject's cognitive delay segment;

[0033] Step S214: Anonymize the subject's voice and replace personal identification information with a numerical identifier;

[0034] In step S22, during the subject's described speech feature extraction step, the openSMILE toolkit is used to extract 1582-dimensional acoustic features defined by the INTERSPEECH 2010 feature set standard, covering fundamental frequency, energy distribution, spectral characteristics, and Mel frequency cepstral coefficient parameters.

[0035] In step S23, the deep learning model is a long short-term memory network.

[0036] A voice-based depression early warning system employs a voice-based intelligent depression recognition method, comprising:

[0037] The data acquisition module is configured to sequentially present standardized Rorschach inkblot test cards to the subjects, record the subjects' spoken descriptions, and preprocess the raw speech.

[0038] The feature extraction module is configured to extract acoustic features from the preprocessed speech.

[0039] The model calculation module is configured to input the extracted acoustic features into a pre-trained deep learning model to calculate a depression risk score.

[0040] The results output module is configured to output a depression risk assessment result based on the risk score P.

[0041] In the data acquisition and preprocessing steps, the original speech is subjected to noise reduction, segmentation, and format standardization preprocessing.

[0042] In the subject-described speech feature extraction step, the extraction of the 1582-dimensional acoustic features is based on the preprocessed speech data. The openSMILE toolkit is used to automatically extract features according to the INTERSPEECH 2010 feature set standard, thereby determining 1582-dimensional acoustic features, covering parameters such as fundamental frequency (F0), energy distribution, spectral characteristics, and Mel frequency cepstral coefficients (MFCC).

[0043] In the model recognition step, the LSTM network automatically learns depression-related dynamic patterns in speech features through its temporal modeling capabilities, and derives a risk score P.

[0044] If the risk score P is greater than or equal to the preset threshold of 0.5, it is determined to be at high risk of depression.

[0045] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements a speech-based intelligent method for recognizing depression.

[0046] The technical features not described in detail in this solution are based on the conventional understanding of those skilled in the art and can be implemented in conjunction with existing technologies, and will not be elaborated further here.

[0047] Compared with existing technologies, the depression intelligent recognition scheme based on Rorschach inkblot test speech provided by this invention has the following significant technical effects:

[0048] (1) Improved objectivity of recognition results: This invention extracts 1582 objective acoustic features by collecting the descriptive speech of subjects when describing pictures, replacing the method that relies entirely on self-report, and can effectively bypass the psychological defense mechanisms of subjects. Speech feature extraction is less affected by social expectations or cognitive control, and can directly reflect emotional state and cognitive patterns. This greatly reduces the interference of subjective factors such as social desirability and concealment behavior on the evaluation results, thereby obtaining physiological and psychological data with a higher signal-to-noise ratio and greater authenticity and reliability, fundamentally overcoming the defects of self-report scales that are highly subjective and easy to falsify.

[0049] (2) Optimized assessment process: This invention constructs a complete process that includes voice data acquisition, openSMILE automated feature extraction, LSTM deep learning model calculation, and result output. This replaces the traditional Rorschach test coding and scoring process, which relies entirely on expert experience and takes several hours, with automated processing and deep learning recognition of the raw Rorschach test voice data. Depression risk assessment results can be output without manual coding. The system can quickly complete feature extraction and model calculation, generating a depression risk assessment report in a short time. Compared to purely manual coding analysis, it reduces the high dependence on expert experience, making rapid and universal psychological screening of large-scale groups possible.

[0050] (3) Early warning is achieved: The present invention integrates a risk assessment function, which can determine the risk of depression based on the depression risk score calculated by the LSTM model. This enables mental health service workers to identify high-risk individuals in a timely manner, rather than passively waiting for problems to occur before taking action, which significantly improves the effectiveness of mental health crisis intervention.

[0051] (4) Scalability: The core of this invention lies in a deep learning model and system architecture based on Rorschach test speech, and its underlying technical solution is replicable. This system can not only be applied to depression screening for specific occupational groups, but also, after adjustment, can be extended to psychological health screening scenarios such as occupational entry psychological assessment, pre-employment psychological health re-examination, psychological services during job transfer and resignation, and psychological assessment of students in related professional colleges.

[0052] In summary, this invention organically combines the Rorschach inkblot test with deep learning speech analysis technology to create a speech-based intelligent depression recognition method. This method effectively solves the problems of "high subjectivity" and "low efficiency" in the current field of depression recognition and has high application value. Attached Figure Description

[0053] Figure 1 This is a system architecture diagram of Embodiment 1 of the present invention.

[0054] Figure 2This is a flowchart of voice data preprocessing in Embodiment 1 of the present invention.

[0055] Figure 3 A schematic diagram of the deep learning model training process based on Rorschach inkblots in Example 1. Detailed Implementation

[0056] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0057] Example 1

[0058] See Figures 1-3 A voice-based intelligent method for recognizing depression includes the following steps:

[0059] Step S1: Model training step. Based on the Rorschach inkblot test, an RTSC-DEP speech database is constructed. Clinical depression diagnosis results are used as supervision signals to train the deep learning model and obtain the depression recognition model.

[0060] Initial participants were selected through a volunteer recruitment process. First, a recruitment notice was distributed, containing a questionnaire adapted from the Self-Rating Depression Scale (SDS). Basic demographic information such as gender and age was collected, and participants were required to provide their real contact information, but their real names were not required. To avoid suggestive language, the recruitment notice avoided the words "depression" and related terms, using softening keywords such as "low mood," "loss of interest," and "desire for healing" to guide volunteers in self-assessment, ensuring that the self-rating scale accurately reflected the participants' psychological state.

[0061] All participants were volunteers with the aim of gaining self-awareness of their psychological state, effectively avoiding data contamination caused by factors such as feigning illness or faking illness. Initial selection of positive and control samples was based on SDS scores. The depression group was selected from highest to lowest SDS scores, while the control group was selected from lowest to highest SDS scores. After contacting volunteers, their consent was obtained, and they signed informed consent forms for online interviews. One-on-one online interviews were scheduled with a chief psychiatrist. The interview process was double-blind, meaning neither the psychiatrist nor the participant had access to the SDS test results. The psychiatrist determined the participants based on the diagnostic criteria for depression in the 11th revision of the International Classification of Diseases (ICD-11).

[0062] Based on the doctor's diagnosis, participants meeting the inclusion criteria were contacted again, and the Rorschach test was administered within one week of the interview. Before the test, participants' consent was obtained for audio and video recording, and they signed an informed consent form for the Rorschach test. This study used a "scale initial screening + doctor's diagnosis" approach to determine enrolled participants, and constructed a multi-dimensional depressive speech corpus (RTSC-DEP) based on the Rorschach Inkblot Comprehensive System (CS) behavioral experiment framework. Core assessment tools included:

[0063] 1. Self-Rating Depression Scale (SDS)

[0064] This scale contains 20 standardized items, comprehensively covering the four dimensions of depressive states: psychomotor symptoms (such as depressed mood, worse during the day and better at night), somatic manifestations (such as fatigue and sleep disturbances), psychomotor symptoms (such as slowed thinking or agitation), and cognitive disturbances (such as feelings of hopelessness and irritability). This multidimensional structural design effectively reflects the clinical heterogeneity of depression. In terms of measurement characteristics, the SDS employs a four-point scoring method and a balanced scoring strategy, that is, 10 positive and 10 negative statements are alternately arranged. For example, the paired setting of "depressed mood" (positive scoring) and "good sleep" (negative scoring) reduces the threat of response bias to validity.

[0065] 2. International Classification of Diseases, Eleventh Revision (ICD-11)

[0066] ICD-11 is the latest global classification standard for diseases officially released by the World Health Organization (WHO) in 2019 and effective in January 2022. It aims to help doctors more accurately match disease definitions and reduce misdiagnosis and missed diagnosis. It provides a unified scientific framework for the diagnosis, statistics, and management of diseases, disabilities, injuries, and health states.

[0067] Step S2: Model Application Steps. The Rorschach inkblot test cards are presented to the subject, voice recordings are collected and input into the trained depression recognition model, and the depression risk assessment results are output. This process mainly consists of the following four steps:

[0068] Step S21: Data Acquisition and Preprocessing. The Rorschach inkblot test cards are presented to the participants, and their spoken descriptions are recorded. The raw speech is then preprocessed with noise reduction, segmentation, and format standardization. Ten standardized Rorschach inkblot test cards are used as stimulus materials. The testing process is conducted by an examiner who has received systematic Rorschach training, strictly adhering to the Exner integrated system specifications.

[0069] This invention employs a four-level preprocessing workflow to clean and standardize the raw speech data. For details, please refer to [link / reference needed]. Figure 2 .

[0070] The recording device was a high-fidelity directional microphone (MAILADA S1000). To improve audio quality and eliminate equipment differences, the experimenter pre-set the high-fidelity directional mode (signal-to-noise ratio ≥70dB), setting the sampling depth to 64-bit and the sampling rate to 48kHz. The audio was saved in mono WAV lossless format. Next, automatic low-frequency noise processing was performed. Adobe Audition's AI noise reduction algorithm based on spectrum analysis can identify and suppress steady-state low-frequency noise. The initial value was set to -12dB in the "Noise Reduction" function, with a cutoff frequency of 200Hz. The AI ​​algorithm can intelligently distinguish between low-frequency environmental noise and human speech, preserving mid-to-high frequency vocal details. All the above data are within the experimental error range and are acceptable. Based on this, people can determine the patterns of data variation and draw relevant conclusions. The third step is the manual processing stage of voice editing. Each stage of the test contains a large amount of the examiner's voice. Since the examiner's voice information is irrelevant to the intelligent recognition task of this study, it needs to be removed. Subjects' spontaneous expression and silent thinking contain unconscious emotional leaks (such as fundamental frequency fluctuations and speech rate changes), which are key biomarkers for depression identification. To ensure the effective use of natural silent segments in the voice data, selective retention of recording segments is necessary based on the cognitive activity information content. Sudden high-frequency noise is a non-silent interference; while Audition's noise reduction tool can reduce background noise, its ability to handle specific sudden noise is limited, requiring manual deletion. The specific operational standards for voice editing are as follows:

[0071] 1. Retained Part

[0072] (1) Subject’s active expression segment: The complete statement from the beginning to the end of the subject’s speech (e.g., “It looks like a butterfly with white patterns on its wings…”) is fully preserved to capture the emotional projection content and the acoustic features of the speech.

[0073] (2) Subject cognitive delay: The time interval between the experimenter finishing speaking and the subject starting to speak (such as the subject's silence period after the experimenter asks a question). Such segments are retained to analyze non-verbal cognitive information such as response delay and breathing rhythm.

[0074] 2. Elimination of certain parts

[0075] (1) Intervention segment by the experimenter: The period from when the subject finishes speaking to when the experimenter finishes speaking (such as when the experimenter asks follow-up questions or provides explanations). Since this includes the experimenter's thinking and guiding language, it should be removed to avoid interference.

[0076] (2) Closed-ended question and answer segment: a short confirmatory dialogue during the questioning phase (e.g., experimenter: "Is this the place?" subject: "Yes."). Such segments are eliminated because they lack deep cognitive information and have simple acoustic features.

[0077] (3) Other high-frequency noises: including sudden plosive sounds (such as sudden environmental noise) and sneezing, yawning, throat clearing, and persistent coughing. These can be marked and removed by observing abnormal changes in speech amplitude. If the interference such as coughing sounds is similar to the effective speech in frequency, more time is needed to judge and remove it.

[0078] The retained segments are arranged in their original chronological order after editing.

[0079] 3. Anonymization of participant identity

[0080] Audio files are assigned digital numbers to remove personal identification information.

[0081] The cleaned and edited audio data reduces irrelevant interference and low-information segments, improving data privacy and purity, and laying the foundation for intelligent analysis of Rorschach audio.

[0082] Step S22: Feature extraction step. The openSMILE toolkit is used to extract 1582-dimensional acoustic features defined by the INTERSPEECH 2010 feature set standard from the preprocessed speech. The features include at least the fundamental frequency and its coefficient of variation, Mel frequency cepstral coefficients, energy distribution parameters, speech rate and pause frequency.

[0083] In this step, feature extraction utilizes openSMILE, an open-source toolkit for speech emotion feature extraction, to extract 1582-dimensional acoustic features defined by the INTERSPEECH 2010 feature set (IS_10) standard. This feature set consists of three parts: based on 68 fundamental low-level descriptors (LLDs) and their corresponding dynamic incremental parameters (Delta coefficients), these 68 parameters are expanded in multiple dimensions using 21 higher-order statistical functions (such as mean, standard deviation, kurtosis, skewness, percentiles, interquartile range, etc.), generating 1428 statistically derived features; for the 8 LLDs and their Delta coefficients related to the fundamental frequency of speech, 152 pitch-related features are extracted using 19 specialized statistical functions. Furthermore, the system simultaneously records two time-series indicators: the number of pseudo-syllables and the total duration of the speech stream, forming a comprehensive parameter system with a total of 1582 dimensions of features.

[0084] For each participant, assuming there are T speech segments, openSMILE will output a two-dimensional feature matrix of shape (T, 1582), where each row corresponds to a 1582-dimensional acoustic feature at one time step, and the row order strictly matches the temporal order of the speech segments. The feature data from all N participants constitute a three-dimensional tensor of shape (N, T, 1582), which serves as the direct input to the LSTM (Long Short-Term Memory) model.

[0085] Step S23: Model recognition step, after standardization of the 1582-dimensional speech features, input into the LSTM deep learning model, calculate the risk score P of the subject being identified as having a depressive state, and output the depression risk assessment result based on the risk score P.

[0086] In this step, the 1582-dimensional features are first standardized using Z-score:

[0087] in The mean of this feature for the training set, The standard deviation of this feature in the training set is given. After standardization, the feature matrix of each subject has the shape (T, 1582), where T is the number of speech segments for that subject.

[0088] The constructed LSTM network structure is as follows: the first LSTM layer contains 128 memory units; the second LSTM layer contains 64 memory units; followed by a fully connected layer with 32 units and the activation function is ReLU; the output layer has one neuron with the activation function sigmoid, outputting the depression probability P. The model uses binary cross-entropy as the loss function, and the Adam optimizer learning rate is set to 0.001.

[0089] The structure of neurons in the LSTM network includes cell state C. t Forgotten Gate t Input gate i t and output gate o t The cell state is updated through three gate structures. The calculation formulas for different gates in a neuron cell are as follows:

[0090]

[0091] Among them, f t i t o t C t , W represents the candidate states of the forget gate, input gate, output gate, and the output and input of the memory cell, respectively. f W i W c W o b f b i b c b o These represent the corresponding weight coefficient matrix and bias term, respectively. and tanh represent the sigmoid and hyperbolic tangent activation functions, respectively.

[0092] Step S24 Result Output Step: Based on the risk score P, output the depression risk assessment result.

[0093] Risk score for depression The calculation formula is as follows:

[0094]

[0095] in, This represents the hidden state output at the last time step of the second-layer LSTM. This represents a fully connected function. The Sigmoid function is represented by the following formula:

[0096]

[0097] The model outputs a depression risk score P between 0 and 1. If P ≥ 0.5, it is considered a high risk of depression; if P < 0.5, it is considered a low risk of depression. This threshold can be adjusted according to the sensitivity and specificity requirements of the actual application scenario. In this embodiment, 0.5 is used as the default threshold.

[0098] Suppose a subject's speech is processed using OpenSMILE to extract 1582-dimensional features, which are then standardized and input into a trained LSTM model. The model calculates the subject's probability of depression as P=0.87. This result indicates that the subject has a high risk of depression (87%), and the system classifies them as having a high risk of depression.

[0099] The model was validated using the scheme in this embodiment, and the results are as follows: accuracy 0.826, recall 0.917, and F1 score 0.846.

[0100] Detailed explanation of step S22: In this embodiment, the acoustic feature extraction step based on the preprocessed speech data does not require manual intervention. The openSMILE toolkit extracts 1582-dimensional acoustic features independently for each segmented speech segment according to the preset IS_10 feature set standard.

[0101] The naming of all 1582 features follows the OpenSMILE naming convention. For example, according to the OpenSMILE naming convention, each part of the feature name represents a different processing step and attribute. For example, "logMelFreqBand" refers to the Mel band energy after logarithmic processing, "sma" represents Simple Moving Average, "de" represents Delta energy, and "pctlrange", "linregerrQ", "percentile", "iqr", and "stddev" are different statistics.

[0102] Example 2

[0103] A voice-based depression early warning system includes:

[0104] The data acquisition module is configured to present 10 standardized cards of the Rorschach inkblot test to the subjects in sequence, record the subjects' spoken descriptions, and perform noise reduction, segmentation, and format standardization preprocessing on the original speech.

[0105] The feature extraction module is configured to extract 1582-dimensional acoustic features defined by the INTERSPEECH 2010 feature set standard from preprocessed speech using the openSMILE toolkit. The features include at least the fundamental frequency and its coefficient of variation, Mel frequency cepstral coefficients, energy distribution parameters, speech rate and pause frequency parameters.

[0106] The model recognition module is configured to input standardized 1582-dimensional speech features into an LSTM deep learning model to calculate a risk score P for the subject's depressive state. The structure of the neurons in the LSTM network includes cell states C. t Forgotten Gate t Input gate i t and output gate o t ;

[0107] The results output module is configured to output a depression risk assessment result based on the risk score P.

[0108] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a depression intelligent early warning method based on Rorschach inkblot test speech.

[0109] The system architecture implementation process is as follows:

[0110] Although the voice data acquisition process requires the experimenter to follow a standardized procedure, the feature extraction, model calculation, and result output stages of this invention are entirely automated by computer. Specifically, a B / S architecture system can be constructed, including:

[0111] Front-end interface: Serving as the interaction point between the system and operators, this interface allows experimenters to input basic information about participants and supports voice file uploads. Uploaded voice files will be sent to the back-end for automated processing. For steps requiring manual intervention, such as manual review in special circumstances, the front-end provides a corresponding user interface.

[0112] Backend service: Receives the audio file uploaded from the frontend, uses the openSMILE toolkit to automatically extract 1582-dimensional acoustic features, and performs Z-score normalization on the features. Then, the normalized feature matrix is ​​input into a pre-trained LSTM deep learning model for calculation, obtaining the depression risk score P. The backend service is responsible for calling stored model parameters, such as LSTM network weights and bias terms, to calculate the risk assessment P value.

[0113] Database: Stores participants' basic information, original audio files, extracted feature data, model parameters, model calculation results, and system operation logs. The database uses encrypted storage to ensure the sensitivity and privacy of the psychological data.

[0114] Results Output Module: After the model calculation is complete, a structured depression risk assessment report is automatically generated, mainly including depression risk score information. The report can be exported in PDF or HTML format for reference by mental health service workers.

[0115] Early warning module: When the depression risk score P calculated by the model is ≥0.5, the system determines that the risk of depression is high. This threshold can be adjusted according to the sensitivity and specificity requirements of the actual application scenario. In this embodiment, 0.5 is used as the default judgment threshold.

[0116] The technical details not described in this solution are all based on the conventional understanding and operation of those skilled in the art and can be accomplished using existing technologies, and will not be elaborated here.

[0117] Example 3

[0118] The core difference between this embodiment 3 and the aforementioned embodiment 1 is that: only the speech data from the free-response phase is used for model training and recognition, without collecting speech information from the questioning phase, thus constructing a simplified speech database and recognition model. All other aspects not described herein are the same as in embodiment 1 and will not be repeated here.

[0119] Step S1: Model Training Steps

[0120] In this embodiment, only the speech data from the free response phase is used for model training during the speech database construction phase, without collecting speech data from the questioning phase, thus constructing a simplified depression speech database (RTSC-DEP-Lite1).

[0121] The specific operations for subject screening and enrollment, voice data collection and preprocessing, and data annotation are the same as in Example 1. The difference is that only the voice data from the free-response phase is used for model training. The LSTM deep learning model is trained using clinical depression diagnosis results as a supervisory signal to obtain a depression recognition model.

[0122] Step S2: Model Application Steps

[0123] All Rorschach inkblot test cards are presented to the participants sequentially. During the free-response phase, the participants' spoken descriptions are collected. After preprocessing and feature extraction, the audio is input into a trained LSTM model to calculate a depression risk score P. The risk assessment result is output based on the P value. If P ≥ 0.5, the participant is considered to have a high risk of depression; if P < 0.5, the participant is considered to have a low risk of depression. This threshold can be adjusted according to the sensitivity / specificity requirements of the actual application scenario. In this embodiment, 0.5 is used as the default threshold.

[0124] The model was validated using this embodiment, and the results are as follows: accuracy 0.870, recall 0.889, and F1 score 0.879.

[0125] The other parts of this embodiment (including system architecture, computer-readable storage medium, B / S architecture implementation, etc.) are exactly the same as those in embodiment 1, and will not be described again here.

[0126] Example 4

[0127] The core difference between this Example 4 and the aforementioned Example 1 is that, starting from the model training phase, only a subset of the cards (cards VIII, IX, and X) are selected as stimulus materials, and only speech information from the free-response phase is collected to construct a simplified speech database and recognition model. All other aspects are the same as in Example 1 and will not be repeated here.

[0128] Step S1: Model Training Steps

[0129] In this embodiment, only three color cards from the Rorschach Inkblot Test (cards VIII, IX, and X) are used as stimulus materials during the speech database construction phase to collect speech information during the free response phase and construct a simplified depression speech database (RTSC-DEP-Lite2).

[0130] The specific operations for subject screening and enrollment, voice data collection and preprocessing, and data annotation are the same as in Example 3, except that the stimulus materials are replaced by cards VIII, IX, and X instead of cards I-X, and voice information is not collected during the questioning phase. The LSTM deep learning model is trained using clinical depression diagnosis results as a supervisory signal to obtain a depression recognition model.

[0131] Step S2: Model Application Steps

[0132] Participants were presented with Rorschach inkblot test cards VIII, IX, and X in sequence. Their spoken descriptions during the free-response phase were collected, preprocessed, and feature-extracted before being input into a trained LSTM model. A depression risk score P was calculated, and a risk assessment result was output based on the P value. If P ≥ 0.5, the participant was considered to have a high risk of depression; if P < 0.5, the participant was considered to have a low risk of depression. This threshold can be adjusted according to the sensitivity / specificity requirements of the actual application scenario; in this embodiment, 0.5 is used as the default threshold.

[0133] The model was verified using the scheme of this embodiment, and the results are as follows: the F1 scores of charts VIII, IX, and X are 0.870, 0.909, and 0.833, respectively.

[0134] The other parts of this embodiment (including system architecture, computer-readable storage medium, B / S architecture implementation, etc.) are exactly the same as those in embodiment 1, and will not be described again here.

[0135] This invention combines the Rorschach inkblot test with deep learning speech analysis technology, which, while ensuring the depth of psychological assessment, achieves efficient, objective, automated intelligent identification and early warning of depressive states. It significantly reduces the reliance of traditional methods on expert experience and improves the feasibility and efficiency of large-scale psychological screening.

[0136] Example 5

[0137] In this embodiment 5, the present application provides a computer-readable storage medium that stores computer instructions. When a computer reads the computer instructions in the storage medium, the computer executes the aforementioned voice-based intelligent depression recognition method and intelligent early warning system.

[0138] It is understood that the voice-based intelligent depression early warning system described in the above embodiments can be implemented by a computer chip or physical entity, or by a product with certain functions. The specific form of the computer described in this solution can be a personal computer, a personal digital assistant, a tablet computer, a laptop computer, a wearable device, or any combination of these devices, etc.

[0139] The computer storage media described in this solution are not limited to phase-change memory, other types of random access memory, read-only memory, static random access memory, dynamic random access memory, electrically erasable programmable read-only memory, digital multifunction optical disc, magnetic tape, disk storage, quantum memory, flash memory, read-only optical disc, graphene-based storage media, magnetic storage devices, and non-transfer media, and can be used to store information that can be accessed by computing devices.

[0140] It should be noted that the user information (including but not limited to device information, personal information, etc.) and data (including but not limited to data analysis, data storage, data display, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and can comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0141] Additional Explanation: In this invention, the "intelligent recognition method" and the "intelligent early warning system" constitute a complete technical solution, with complementary and progressively related functions. The "intelligent recognition method" focuses on the technical implementation level, that is, through the processing and modeling of speech signals, it outputs a quantitative indicator (risk score P) reflecting the degree of depression risk. Here, "recognition" refers to the detection and quantification of depression-related patterns in speech signals, rather than making a clinical diagnosis of whether an individual has depression. The "intelligent early warning system" focuses on the functional application level, that is, based on the comparison result of the risk score P and a preset threshold, it automatically triggers an early warning notification to prompt relevant personnel to pay attention to potential risks. Here, "early warning" refers to information push and risk alerts, and does not replace the professional judgment of clinicians. Together, they constitute the overall technical solution for auxiliary assessment and early warning of depression risk.

[0142] The technical features of this invention not described can be implemented by or using existing technology, and will not be repeated here. Of course, the above description is not a limitation of this invention, and this invention is not limited to the examples above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention should also be within the protection scope of this invention.

Claims

1. A speech-based intelligent method for recognizing depression, characterized in that, Includes the following steps: Step S1: Model training step. Based on the Rorschach inkblot test, an RTSC-DEP speech library is constructed. The deep learning model is trained using depression diagnostic labels as supervision signals to obtain a depression recognition model. Step S2: Model application step. Present the Rorschach inkblot test cards to the subject, collect voice recordings and input them into the trained depression recognition model, and output the depression risk assessment results.

2. The speech-based intelligent depression recognition method according to claim 1, characterized in that, In step S1, the method for constructing RTSC-DEP includes the following steps: Step S11: Subject screening and enrollment. Initial screening was conducted using the Self-Rating Depression Scale, combined with structured clinical interviews with psychiatrists, and participants were divided into a depression group and a control group. Step S12: Voice data collection. The standardized Rorschach inkblot test cards are presented to the participants, and the participants' descriptive voices are collected and recorded throughout the process. Step S13: Speech preprocessing, which involves noise reduction, format standardization, and removal of invalid segments from the collected raw speech; Step S14: Data labeling, labeling the voice data based on the structured clinical interview diagnostic results of ICD-11.

3. The speech-based intelligent depression recognition method according to claim 2, characterized in that, In step S11, the method of "initial screening by scale + doctor's diagnosis" is used to determine the subjects to be included in the group. The core assessment tools include the Depression Self-Rating Scale and ICD-11.

4. The speech-based intelligent depression recognition method according to claim 2, characterized in that, In step S14, the depression group and the control group are strictly divided based on the results of the structured clinical interviews conducted by ICD-11.

5. The speech-based intelligent depression recognition method according to claim 1, characterized in that, Step S2 specifically includes the following steps: Step S21: Data collection and preprocessing step. Present the Rorschach inkblot test cards to the subjects, record the subjects' descriptive speech, and preprocess the raw speech. Step S22: Feature extraction step, extracting a set of acoustic features that can represent emotions, cognition and psychomotor states from the preprocessed speech; Step S23: Model recognition step, after the extracted acoustic features are standardized, they are input into the pre-trained deep learning model to calculate the risk score P of the subject belonging to the depressive state; Step S24: Result output step, based on the risk score P, output the depression risk assessment result. Specifically, step S21, which involves preprocessing the original speech by noise reduction, segmentation, and format standardization, includes the following steps: Step S211: Record using a high-fidelity directional microphone and save as a mono audio file in lossless audio format; Step S212: Use a noise reduction algorithm based on spectrum analysis to identify and suppress steady-state low-frequency noise, and adaptively adjust the noise reduction parameters according to the ambient noise level; Step S213: Manually edit out the experimenter's intervention segment, closed-ended question and answer segment and other high-frequency noise, and retain the subject's active expression segment and the subject's cognitive delay segment; Step S214: Anonymize the subject's voice and replace personal identification information with a numerical code.

6. The speech-based intelligent depression recognition method according to claim 5, characterized in that, In step S22, during the subject's described speech feature extraction step, the openSMILE toolkit is used to extract 1582-dimensional acoustic features defined by the INTERSPEECH 2010 feature set standard, covering fundamental frequency, energy distribution, spectral characteristics, and Mel frequency cepstral coefficient parameters.

7. The speech-based intelligent depression recognition method according to claim 5, characterized in that, In step S23, the deep learning model is a long short-term memory network.

8. A voice-based depression early warning system, employing a voice-based intelligent depression recognition method as described in any one of claims 1-7, characterized in that, include: The data acquisition module is configured to sequentially present standardized Rorschach inkblot test cards to the subjects, record the subjects' spoken descriptions, and preprocess the raw speech. The feature extraction module is configured to extract acoustic features from the preprocessed speech. The model calculation module is configured to input the extracted acoustic features into a pre-trained deep learning model to calculate a depression risk score. The results output module is configured to output a depression risk assessment result based on the risk score P.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images