Systems and methods for message generation

Abstract

A message can be generated for a patient using a message template, a machine-learning model, and recommendation model. A parameterized version of the message template can be generated using the recommendation model and a profile of the patient. The machine-learning model can be configured to restate the parameterized version of the message using the patient profile and a supervisor profile. The restated message can be delivered to the patient or approved by a supervisor and then delivered to the patient. Based on supervisor feedback concerning the message template, the parameterized message template, or the restated message template, the recommendation model or the machine-learning model can be trained or fine-tuned.

Description

SYSTEMS AND METHODS FOR MESSAGE GENERATIONCROSS REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Application No. 63 / 509,667, filed June 22, 2023, the content of which are incorporated herein by reference in its entirety.BACKGROUND

[0002] Patient-support platforms have the potential to improve patient mental and physical health. But when a patient-support platform fails to engage a patient, this potential remains unrealized. Patient-support platforms can improve patient engagement by providing relevant, personalized content. Such content can include feedback regarding the patient or patient interactions with the patient-support platform. However, personalized feedback can be difficult to provide in a controlled, scalable manner.SUMMARY

[0003] The disclosed systems and methods relate to an architecture for providing personalized messages in a controlled, scalable manner.

[0004] The disclosed embodiments include a message-generation system. The messagegeneration system can include at least one processor and at least one computer-readable medium containing instructions. When executed by the at least one processor, the instructions can cause the message-generation system to perform operations. The operations can include, in accordance with a schedule or in response to an event associated with a user, generating a parameterized message using a message template, generating a restated message by applying the parameterized message to a machine-learning model configured to restate the parameterized message, and providing the restated message to the user.

[0005] The disclosed embodiments include another message-generation system. The message-generation system can include at least one processor and at least one computer-readable medium containing instructions. When executed by the at least one processor, the instructions can cause the message-generation system to perform operations. The operations can include generating a restated message using a machine-learning model, a recommendation model, a message template, a profile of a patient, and a supervisor profile. The operations can further include displaying, in a first view of a supervisor interface, a communication including the restated message or multiple selectable restated messages including the restated message. The operations can further, in response to an interaction of a supervisor with the supervisor interface, providing the restated message to the patient.

[0006] The disclose embodiments further include methods and non-transitory, computer- readable media corresponding to the above systems.

[0007] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosed embodiments, as claimed.BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The accompanying drawings, which comprise a part of this specification, illustrate several embodiments and, together with the description, serve to explain the principles and features of the disclosed embodiments. In the drawings:

[0009] FIG. 1 depicts an exemplary architecture for providing a patient-support platform, consistent with disclosed embodiments.

[0010] FIG. 2 depicts exemplary contents of a data storage of the architecture of FIG. 1, consistent with disclosed embodiments.

[0011] FIG. 3 depicts exemplary contents of a message bank of the data storage of FIG. 2, consistent with disclosed embodiments.

[0012] FIG. 4 depicts a schematic of an exemplary system for generating messages, consistent with disclosed embodiments.

[0013] FIG. 5 A depicts an exemplary communication containing a set of message templates, consistent with disclosed embodiments. FIGs. 5B to 5E depict exemplary restatements of parameterized messages generated from the message templates depicted in FIG. 5A.

[0014] FIG. 6 A depicts an exemplary communication containing a second set of message templates. FIG. 6B and 6C depict exemplary restatements of parameterized messages generated from the message templates depicted in FIG. 6A.

[0015] FIG. 7 depicts an exemplary method for selecting message templates, consistent with disclosed embodiments.

[0016] FIG. 8 depicts an exemplary method for generating and providing messages to patients, consistent with disclosed embodiments.

[0017] FIG. 9 depicts an exemplary supervisor interface for reviewing messages, consistent with disclosed embodiments.

[0018] FIGs. 10A and 10B depict an exemplary supervisor interface for reviewing messages, in accordance with disclosed embodiments.

[0019] FIG. 11 depicts an exemplary supervisor interface for arranging message templates, in accordance with disclosed embodiments.DETAILED DESCRIPTION

[0020] Reference will now be made in detail to exemplary embodiments, discussed with regard to the accompanying drawings. In some instances, the same reference numbers will be used throughout the drawings and the following description to refer to the same or like parts. Unless otherwise defined, technical or scientific terms have the meaning commonly understood by one of ordinary skill in the art. The disclosed embodiments are described in sufficient detail to enable those skilled in the art to practice the disclosed embodiments. It is to be understood that other embodiments may be utilized and that changes may be madewithout departing from the scope of the disclosed embodiments. Thus, the materials, methods, and examples are illustrative only and are not intended to be necessarily limiting.

[0021] Patient-support platforms have the potential to improve patient mental and physical health. Such platforms can educate patients about their conditions, provide a forum for social interactions and support, and promote beneficial behavior modifications. But when a patientsupport platform fails to engage a patient, this potential remains unrealized. A patient that does not access a platform cannot be educated, socially supported, or trained through that platform.

[0022] Patient-support platforms can improve patient engagement by providing relevant, personalized content, such as personalized encouragement. But such content can be difficult to provide at scale, as different patients have different needs. Content appropriate for one patient may appear inappropriate, patronizing, or irrelevant to another patient.

[0023] United States Patent Application No. 18 / 694,400, incorporated herein by reference in its entirety, discloses patient-support platforms that improve patient engagement through scalable personalization. The disclosed patient-support platforms can provide interaction opportunities to patients. The interaction opportunities can be selected and customized based on obtained patient medical information, prior patient interactions with the patient-support platform, or the like. Patients can interact with the disclosed patient-support platforms through graphical user interfaces configured to provide relevant, personalized content. Patient medical information and patient interactions can be stored in a patient profile.

[0024] The disclosed patient-support platforms further provide supervisory (e.g., coaching, management, teaching, reviewing, or the like) functionality. Such functionality can address the lack of personal contact that afflicts conventional patient-support platforms. Consistent with disclosed embodiments, the patient-support platform user interface can include a supervisory portal that enables supervisors (e.g., coaches, managers, teachers, reviewers, orthe like) to interact with one or more patients. The supervisory portal can support improved scalability by enabling a supervisor to simultaneously (e.g., in a single view) review the statuses and actions of multiple patients.

[0025] In some embodiments, the supervision portal can support interactions between supervisors and patients. The interactions can be designed to further treatment and prevent patient disengagement. United States Patent Application no. 18 / 694,400 particularly discloses that supervisors can provide messages to patients. Such messages can improve the patient experience and support greater patient engagement with the patient-support platform.

[0026] As may be appreciated, conventional patient-support platforms can struggle with generating personalized, relevant feedback messages for patients in a controllable, scalable manner. Approaches that rely on supervisors to draft each feedback message to each patient may not be scalable. Furthermore, supervisors may vary in quality, training, and commitment, with some supervisors providing superior feedback to others. In contrast, approaches that automatically send standard messages may encourage patient disengagement. Patients may quickly determine that standard messages are sent automatically, as such messages may be repetitive or characterized by grammatical or stylistic mistakes or infelicities. The disjunction between the individualized, supportive form of standard messages and the automated means by which they are generated may breed annoyance, alienation, or cynicism in patients, reducing patient engagement with the patient-support platform.

[0027] Emerging machine-learning models, such as large language models, can generate messages in response to prompts. Multi-modal, machine-learning models can receive prompts in multiple modes (e.g., text, audio, images) and may similarly provide messages including content in multiple modes. Such messages are generated in a non-deterministic manner. An approach to generating automatic messages that uses such machine-learning models mayproduce messages that are less repetitive or exhibit fewer grammatical or stylistic mistakes or infelicities than an approach that uses standard messages.

[0028] However, approaches using machine-learning models can exhibit other deficiencies. In particular, such machine-learning models may generate incorrect or offensive content. For example, a machine-learning model may suggest unhealthful crash dieting to a patient seeking help with an overeating disorder or may inaptly suggest glue as a pizza topping to a patient seeking healthy recipes. Furthermore, such machine-learning models largely function as black-boxes, making debugging or preventing such behavior difficult. As the machinelearning model may be provided as a service by a third party, accountability may also be a problem when the machine-learning model provides incorrect or offensive content.

[0029] The disclosed embodiments enable a patient-support platform to provide relevant, personalized messages to patients in a controlled, scalable manner. Consistent with disclosed embodiments, the patient-support platform can include a message bank and recommendation model(s). The message bank can include message templates. The patient-support platform can be configured to recommend message templates using the recommendation model(s) and a patient profile. The recommendation process can be multi-phase, with an initial rule-based phase and a subsequent recommendation phase. In some embodiments, recommendation model(s) can select or rank message templates using category associations or embeddings for the message templates. Thus, a message template may not be tied to a particular program or usage case. Instead, the message template may be recommended for similar patients or in similar situations. The recommendation model(s) may be trained to recommend the message over a broad range of situations when the message proves broadly appropriate and effective. Semantic filtering may also be performed to prevent recommendation of excessively similar messages. The patient-support platform can restate the recommended messages using a machine-learning model, the patient profile, and optionally a supervisor profile. Acommunication including one or more restated messages can then be provided to the patient. In this manner, the message template can control message content, the supervisor can control message sentiment, and the machine-learning model can provide suitable variation in message details.

[0030] The disclosed embodiments can be configured to maintain human control over messages sent to patients. A supervisor may review recommended message templates (or parameterized versions thereof, or the like). The supervisor may select message templates (or parameterized versions thereof, or the like) for restatement. Similarly, a supervisor may review restated messages (or a communication containing restated messages). The supervisor may select, rearrange, or edit the restated messages. In this manner, a supervisor can retain responsibility for any messages sent to the patient.

[0031] The disclosed embodiments can be configured to improve over time. In some embodiments, supervisor feedback can be used to train recommendation models. For example, when a supervisor selects a message template (or a parameterized or restated message generated using the message template), the patient platform may become more likely to recommend the message template. In addition, patient feedback regarding messages, whether express or implicit (e.g., opening or reading messages, following recommendations included in the messages, deleting messages, deleting or ceasing to use the patient-support platform, or the like) can be used to train recommendation models. As an additional example, selection, rearrangement, or editing of restated messages can be used to train or finetune the machine-learning model used to generate the restated messages. In this manner, the patientsupport platform can develop a collection of suitable template messages that satisfy supervisors and drive patient engagement.

[0032] The disclosed embodiments can support additional automation. In some embodiments, message templates can be associated with a confidence value. Supervisor orpatient interactions with the patient support system can affect the confidence value. Interactions that indicate the acceptability or effectiveness of the message template can, for example, increase the confidence value, while interactions that indicate the unacceptability or ineffectiveness of the message template can, for example, reduce the confidence value.Should the confidence value satisfy an automation condition, a message generated using the message template can be sent automatically (or with a certain probability) whenever the message template is selected (e.g., in the first phase of the multi-phase recommendation process). In this manner, the patient-support platform can learn to automatically send acceptable or effective messages to patients, reducing the burden on the supervisors to review such messages.

[0033] As may be appreciated, the disclosed patient support system can improve the scalability of supervisor feedback. Furthermore, the disclosed patient support system can standardize the content and quality of feedback provided by different supervisors. As the machine-learning model can perform translation, the patient support system can be internally language-agnostic: message templates can be stored and parameterized in a common language, then restated into language(s) suitable for supervisor review and / or delivery to patients (e.g., based on indications stored in patient or supervisor profiles).

[0034] Further, although the disclosed embodiments concern a health-oriented, patientsupport platform, the disclosed embodiments are not so limited. Rather, the disclosed embodiments may enable generation of personalized, scalable content in any suitable domain. For example, the disclosed embodiments may assist a teacher in providing feedback to students in an online environment. In another example, the disclosed embodiments may help a teacher review students’ schoolwork and encourage a student to improve their studying habits or continue to submit assignments. Additionally, the disclosed embodiments may assist a customer service manager in feedback sessions with their support team, by discussing thenumber of resolved tickets, average response time, and customer satisfaction ratings in a personal, encouraging manner (e.g., providing personalized feedback to each member of a large support team).

[0035] FIG. 1 depicts an exemplary architecture 100 for providing a patient-support platform, in accordance with disclosed embodiments. Architecture 100 can enable supervisors to provide relevant, personalized messages to patients in a controlled, scalable manner. The relevant, personalized messages can support patient engagement with the platform, which can in turn enhance the therapeutic benefit provided by the platform to patients.

[0036] Architecture 100 can include application system 101 and data storage 102. Applicant system 101 and data storage 102 can be implemented using at least one computing system. In some embodiments, application system 101 and / or data storage 102 can be implemented using a cloud computing system (e.g., Google Could, Amazon AWS, or the like).

[0037] In some embodiments, application system 101 can be implemented as a collection of containerized services. The containerized services can expose endpoints for responding to client requests (e.g., representational state transfer endpoints implemented in Java, or the like). Deployment and management of these services can be achieved using a container orchestration system (e.g., Kubemetes, Docker, Nomad, or the like). In some embodiments, the containers can include, or be configured to connect to volumes that include, data and instructions that specify the patient-support platform.

[0038] As described herein, application system 101 can be configured to obtain and process data and / or instructions from other components of architecture 101 in the performance of the disclosed embodiments. For example, application system 101 can provide parameterized messages to machine learning system(s) 108 for restatement, obtain message templates from data storage 102, provide user interfaces to patient system(s) 103, supervisor system(s) 105,and / or expert system(s) 106, and obtain medical records from patient management system(s) 107, or the like.

[0039] In some embodiments, data storage 102 can be implemented using, in part, an SQL database (e.g., Google Cloud SQL, or the like). The SQL database may be configured to store account information for patients (e.g., patient profile 220, as described herein). In some embodiments, the SQL database may periodically be de-identified and backed up to a data warehouse (e.g., Google BigQuery, or the like). In some embodiments, data storage 102 can be further implemented using, in part, general cloud storage (e.g., Google Cloud Storage, or the like). Such general storage may include storage of files or attachments associated with communications between users of architecture 100.

[0040] In some embodiments, information received from another component can be subject to further processing prior to storing the information in data storage 102. In some embodiments, received images can be subject to image recognition or processing. In some embodiments, structured or unstructured data can be processed to extract predetermined types or categories of information. The processing can be specified by instructions applicable to all patients, particular patients, particular supervisors, particular programs, or the like. For example, a patient on a program that requires meal logging may upload a picture of a meal, together with a description that they were hungry when they started and feel stuffed afterwards. For example, consistent with rules specified for the program, a multi-model, machine-learning model can update a profile of the patient with information indicating that the patient overate and the type of food at the meal. As an additional example, a machinelearning model can be configured to generate structured data describing the physical or mental health of a patient from SOAP notes; medical records; communications from patients, relatives, or caregivers; or the like. In this manner, data storage 102 can be a repository formulti-modal information received from patient systems and other components of architecture100.

[0041] Architecture 100 can include patient system(s) 103, consistent with disclosed embodiments. Such patient systems can include mobile computing devices (e.g., smartphones, tablets, laptops, or the like), or other computing devices (e.g., desktop computers or the like). In some embodiments, the patient systems can be configured to interact with applications system 101 using an application (e.g., a client application native to the particular operating system of the patient system). The application can provide a user interface and handle communications between the patient system and the application system101. In some embodiments, in addition or as an alternative to using the application, patients can access application system 101 through a web portal (e.g., using a web page provided by application system 101, or the like). In some embodiments, the application on the patient system can communicate with application system 101 using HTTPS, or another suitable protocol.

[0042] In some embodiments, patient system(s) 103 can provide to application system 101 patient health information (e.g., number of steps taken, heart rate, pulse oximetry, Electroencephalography (EEG) data, or the like) obtained from patient device(s) 104. In some embodiments, patient system(s) 103 can provide to application system 101 patient interaction information (e.g., patient selections of missions, updates to mission parameters, responses to questionnaires, graphical user interactions, or other patient interaction information, or the like).

[0043] Architecture 100 can include patient devices(s) 104, consistent with disclosed embodiments. In some embodiments, patient devices(s) 104 can include wearable devices, such as smartwatches (e.g., Apple Watch, Samsung Galaxy Watch, or the like), fitness trackers (e.g., Fitbit, or the like), headsets (e.g., MUSE, NeoRythm, or the like), or othersimilar wearable devices. In some embodiments, patient devices(s) 104 can include implantable devices, such as pacemaker / defibrillators; neuro-prostheses for sensory, motorcontrol, pain-management, or other applications; medication pumps; implantable sensors; or the like.

[0044] In some embodiments, as shown in FIG. 1, a patient device can communicate with a patient support system (e.g., with components of architecture 100, such as application system 101), enabling the patient support system to obtain data regarding the patient’s health. In various embodiments, the patient device may be configured to provide this patient health data to a remote system (e.g., an application server associated with the provider of the wearable device). The patient support system can be configured to obtain the patient health data from the remote system. In some embodiments, the wearable device may be configured to provide the patient health data directly to the patient support system (e.g., directly to application system 101, or the like).

[0045] Architecture 100 can include supervisor system(s) 105, consistent with disclosed embodiments. A supervisor system can enable a supervisor to interact with application system 101. In some embodiments, the supervisor can interact with application system 101 to provide relevant, personalized messages to patients in a controlled, scalable manner.

[0046] A supervisor system can be a computing device (e.g., a mobile device, such as a smartphone, tablet, laptop, or the like; or a desktop, terminal, workstation, or the like). In some embodiments, a supervisor system can be configured to interact with applications system 101 using an application (e.g., a client application native to the particular operating system of the supervisor system). The application can be configured to handle communications between the supervisor system and the application system 101. In some embodiments, in addition or as an alternative to using the application, supervisors can access application system 101 through a web portal (e.g., using a web page provided by applicationsystem 101, or the like). In some embodiments, the application on the supervisor system can communicate with application system 101 using HTTPS, or another suitable protocol.

[0047] Architecture 100 can include expert system(s) 106, consistent with disclosed embodiments. An expert system can enable an expert to interact with application system 101. In some instances, the expert can be domain expert. For example, an expert may refer to a medical expert when application system 101 is utilized in a healthcare context. As additional examples, a dietitian may provide messages regarding nutrition, a kinesiologist may provide messages regarding movement management, injury prevention, or exercise improvement, a pain-management doctor may provide messages regarding patient pain, a rheumatologist may provide messages regarding arthritis, or like. In some embodiments, the expert can interact with the expert system to create or modify message templates for use with application system 101, or to create or modify rules for selecting message templates for use in generating messages for a particular patient.

[0048] An expert system can be a computing device (e.g., a mobile device, such as a smartphone, tablet, laptop, or the like; or a desktop, terminal, workstation, or the like). In some embodiments, an expert system can be configured to interact with applications system 101 using an application (e.g., a client application native to the particular operating system of the supervisor system). The application can be configured to provide a user interface and handle communications between the expert system and the application system 101. In some embodiments, in addition or as an alternative to using the application, experts can access application system 101 through a web portal (e.g., using a web page provided by application system 101, or the like). In some embodiments, the application on the expert system can communicate with application system 101 using HTTPS, or another suitable protocol.

[0049] Architecture 100 can include machine learning system(s) 108, consistent with disclosed embodiments. A machine learning system can be a computing device (e.g., amobile device, desktop, terminal, workstation, or the like) or computing system (e.g., a computing cluster, cloud computing platform, or the like) or other suitable system. In some embodiments, the application system 101 can communicate with the machine learning system using HTTPS, or another suitable protocol.

[0050] A machine learning system can be configured to generate personalized messages from parameterized messages using a machine-learning model. In some embodiments, the machine-learning model can be a multimodal model, capable of accepting input and providing output in different modalities (e.g., text, images, audio, or the like). For example, the machine-learning model can be configured to process, analyze, or receive text, images, audio, and / or video as an input. The disclosed embodiments are not limited to any particular type of machine-learning model.

[0051] In some embodiments, the machine-learning model can be or include: a supervised model (e.g., models configured for regression or classification), an unsupervised model (e.g., models configured for clustering), a semi -supervised model, a reinforcement-learning model, a few-shot model, a transfer learning model, and / or a generative model. A generative model can be or include a generative adversarial network (GANs), autoencoder, autoregressive model, diffusion model, transformer-based model, or the like.

[0052] In some embodiments, the machine-learning model can be or include a large language model (LLM). In some embodiments, the machine-learning model can be an artificial intelligence or chatbot model, such as a CHATGPT model (e.g., CHATGPT 3.5, CHATGPT 4, CHATGPT 4o, or the like), a CLAUDE model (e g., CLAUDE 3, or the like), a LLAMA model (e.g., LLAMA 3, or the like), a GEMINI model, or similar artificial intelligence model.

[0053] In some embodiments, the same entity can direct or control application system 101 and machine learning system(s) 108. For example, machine learning system(s) 108 can be orinclude an on-premises machine learning system, such as a machine-learning model running on one or more servers shared with application system 101.

[0054] In some embodiments, components of architecture 100 (e.g., application system 101, or the like) can access machine learning system(s) 108 as a service. For example, machine learning system(s) 108 can host an artificial intelligence model and provide access to the model through an application, application programming interface, or web service.

[0055] Architecture 100 can include patient management system(s) 107, consistent with disclosed embodiments. A patient management system can be a computing device (e.g., a mobile device, desktop, terminal, workstation, or the like) or computing system (e.g., a computing cluster, cloud computing platform, or the like) or other suitable system. In some embodiments, the patient management system can communicate with application system 101 using HTTPS, or another suitable protocol.

[0056] In some embodiments, a patient management system can be configured to maintain patient electronic medical records. The patient management system can be controlled by an entity (e.g., a hospital network, an insurance company, or the like) distinct from the entity controlling application system 101. In some embodiments, there can be a one-direction linkage between the electronic medical records maintained by patient management system(s) 107 and application system 101 : changes made to the electronic medical records can flow to application system 101. In various embodiments, there can be a two-direction linkage between the electronic medical records maintained by patient management system(s) 107 and application system 101 : the changes in one system can flow to the other. Consistent with disclosed embodiments, patient management system(s) 107 can provide application system 101 with patient health information. Application system 101 can use this information to recommend interactive opportunities to the patient as described herein.

[0057] FIG. 2 depicts exemplary contents of data storage 102, in accordance with disclosed embodiments. The depiction in FIG. 2 is a logical depiction: the depicted contents of data storage 102 can be distributed across multiple physical machines (including, in some embodiments, patient system(s) 103, supervisor system(s) 105, and / or patient management system(s) 107, or the like). The depicted contents can include platform components 210, patient profiles 220, and supervisor profiles 230. Platform components 210 can include objects or data useable with multiple patients, while patient profiles 220 can include objects or data associated with a particular patient, and supervisor profiles 230 can include objects or data associated with a particular supervisor. The disclosed embodiments are not limited to any particular implementation or format for platform components 210, patient profiles 220, or supervisor profiles 230.

[0058] Platform components 210 can include message bank 211 and recommendation model(s) 213. Message bank 211 may include message templates. Consistent with disclosed embodiments, message templates can be used to generate controlled, personalized messages that are provided to patients. In various embodiments, the message templates may be generated by an expert (e.g., provided through expert system(s) 106), by a supervisor (e.g., through supervisor system(s) 105), by a machine-learning model of machine learning system(s) 108, or another suitable method.

[0059] In some embodiments, message templates in message bank 211 can be associated with one or more categories. A category can concern a patient condition (e.g., obesity, insomnia, diabetes, or the like), a patient competency (e.g., a particular area of patient capabilities or achievements, such as sleep, stress, activity, nutrition, or psychological health), or a patient interaction. Additionally or alternatively, a category can concern a conversation part. For example, a category can concern greetings (e.g., a greeting from a supervisor to begin a conversation), assignments (e.g., a supervisor giving a task or request toa patient), questions (e.g., a supervisor asking the patient a question), feedback (e.g., a supervisor responding to or giving a critique for a patient action), or the like. For example, a message template can be an assignment (e.g., a conversation part) related to obesity (e.g., a patient condition) and intended to increase the activity level of the patient (e.g., improve a patient competency).

[0060] In some embodiments, as described herein with regards to FIG. 3, a message template can include category values (e.g., numeric or Boolean- valued scalars, vectors, or the like) that indicate the strength of associations between the message template and one or more categories.

[0061] In some embodiments, multiple versions of each message template may be present in message bank 211. In some embodiments, different version may be in different languages. In some embodiments, different versions may correspond to different combinations of characteristics of the patient and the supervisor, such as relative status (e.g., honorific verb forms and vocabulary in Japanese), sex (e.g., as in German and similar languages), familiarity, or the like.

[0062] Recommendation model(s) 213 can be used to recommend message templates in message bank 211. Consistent with disclosed embodiments, the recommended messages can be used to generate personalized messages to patients. Recommendation model(s) 213 are not limited to any particular model architecture and can include rule-based, statistical, or machine-learning models. For example, a recommendation model can select one or more candidate templates based on at least one if-then rule associated with a message template. As an additional example, a feature vector based on patent profiles 220 and / or supervisor profiles 230 can be input to a neural network model. In some embodiments, the neural network model can output a ranking of message templates. A number of the highest ranked candidates can then be selected. As an additional example, a feature vector based on patent profiles 220and / or supervisor profiles 230 can be input to an encoder or embedding model. An output encoding or embedding can be compared to encodings or embeddings generated using message templates. Based on this comparison, a number of message templates can be selected.

[0063] In some embodiments, the recommendation model can use collaborative filtering to recommend selected message templates. Patients and message templates can be associated with embedding vectors (e.g., patients based on patient profiles and embedding vectors based on the text of message templates or categories associations with the message templates). The recommendation model can rank pairings of patients and selected messages. In this manner, for each patient, the recommendation model can determine a ranking of selected messages.

[0064] In some embodiments, a recommendation model can rank message templates for a patient based on category associations for the message templates. As described herein, message templates can have category associations, which may also be expressed using Boolean or numeric scalar or vector valued indicators. The recommendation model can further rank the message templates based on category associations for program(s) associated with the patient and / or patient interactions with the system. For example, a patient may be enrolled in a program for managing diabetes, insomnia, heart disease, or the like. Certain categories can be associated with the program (e.g., patient competencies such as sleep, diet, exercise, or the like). In some embodiments, category associations for the program can be expressed using Boolean or numeric scalar or vector values. Such associations may be automatically determined or manually set. Similarly, patient interactions with the system can be associated with certain categories. For example, food journaling can be associated with a diet patient competency, while completing an exercise mission (e.g., running a mile, taking at least some number of steps, or the like) can be associated with an exercise patient competency. In some embodiments, category associations for the set of user interactions withthe system (e.g., over a period, or a number of recent interactions, or the like) can be expressed using Boolean or numeric scalar or vector values.

[0065] In some embodiments, recommendation model can determine a similarity value for a message template based on category associations for the message template and at least one of category associations for the program(s) or user interactions (or functions thereof). Such a similarity value can be a metric (e.g., an L2 norm, cosine distance, or the like). In determining the similarity value, different categories can contribute different amounts. For example, differing weights can be associated with different categories (or with different Boolean or numeric scalar or vector values that define the category associations). The recommendation model can rank message templates based on these similarity values.

[0066] In some embodiments, the recommendation model can rank message templates based on user interaction data. For example, user retention data or user interaction data can indicate whether a restated message encouraged or discouraged engagement. In some embodiments, when a patient stops using the system (e.g., by deleting the corresponding application from the patient system of the user), a weight can be determined for messages sent to the user. In some embodiments, the weight can be greater for messages sent closer to the date (and optionally time) the user stopped using the system. The recommendation model can use the weight to decrease the ranking of the message template used to generate the restated message.

[0067] In some embodiments, the recommendation model can implement a two-tower embedding model. Such an embedding model can include a patient encoder and a message template, each trained to map from an input feature space to the same encoding space. The two encoders can be trained jointly on the history of patient engagement with messages generated from message templates.

[0068] Patient profiles 220 can include patient information 221 and patient interaction data 223. Patient information 221 can include patient medical information, demographicinformation, location information, or the like. Patient information 221 may also include goals or targets a patient may want to achieve. Patient information 221 can include information obtained from the patient, information obtained from another party regarding the patient (e.g., a parent or guardian, physician, or employer of the patient, or the like), a medical record of the patient (e.g., a medical record accessed or retrieved from patient management system(s) 107), activity data obtained from a device associated with the patient, or the like. In some embodiments, patient information 221 may include information a patient provides to application system 101 through a web portal. Patient information 221 can contain structured or unstructured patient information. For example, patient information can include an assessment of physical fitness on a three-point scale, or an assessment by a physician that the patient is “out of shape.”

[0069] As a first nonlimiting example, patient information 221 may include information for a first user:Name: Jane Sex: F Age: 32 Language: English Activity: High Nutrition: High Sleep: Low Medical History: Recovering from childbirth Goals: Lose weight, Improve Sleep.As indicated by this patient information, Jane is physically fit and eats a balanced diet including fruits and vegetables but has difficulty sleeping regularly and wishes to lose weight. She speaks English as her primary language.

[0070] As a second nonlimiting example, patient information 221 may include information for a second user:Name: JohnSex: MAge: 72Language: EnglishActivity: LowNutrition: LowStress: LowMedical History: Recovering from heart surgeryGoals: Reduce pain from surgery, Improve Activity Level, “be able to play with grandchildren.”As indicated by this patient information, John is not physically fit and does not eat a balanced diet including fruits and vegetables. John seeks to reduce pain from surgery and improve his activity levels. As may be appreciated, the low stress level for John may be determined based on information obtained from multiple differing sources. For example, John’s stress level can be determined based on information included in John’s medical records and obtained from patient management system(s) 107. Such information can include questionnaire results (e.g., from a questionnaire taken by John’s primary care physician), laboratory data (e.g., low or decreasing cortisol levels in laboratory test results), or the like.

[0071] Patient interaction data 223 can include a record of the interactive opportunities selected by the patient, customization options selected by the patient for the selected interactive opportunities, whether the selected interactive opportunities were completed, data pertaining to the performance of the selected interactive opportunities (e.g., steps taken, minutes meditated, questions answered correctly, patient-reported outcome results, patientsubmitted comment text, or the like), communications between the patient and a supervisor, amount of time logged into the platform, or other suitable interactions with the patientsupport platform. For example, patient interaction data 223 may include an indication that the first user is self-reporting lower stress levels, such as the first user submitting the message: “I feel lower stress than last week.”

[0072] In some embodiments, interaction data can be aggregated over a period (e.g., a day, a week, or the like) to generate an indication of the behavior of the user of that period. In such embodiments, the aggregated indication can be used to select message templates according to inclusion / exclusion criteria. In some embodiments, the interaction data can be associated with contextual tags, such as program type or user profile tags.

[0073] Supervisor profiles 230 can include supervisor information 231 and supervisor interaction data 233. Supervisor information 231 can include supervisor demographic information, education or training information, personality information, location information, or the like. Supervisor information 231 may also include information on the style a supervisor may prefer to use with patients (e.g., a coaching style). Supervisor information 231 can include information obtained from the supervisor (e.g., using supervisor system(s) 105), information obtained from another party regarding the supervisor (e.g., an employer of the supervisor), documented information of the supervisor (e.g., a resume accessed or retrieved from supervisor system(s) 105), activity data, or the like. Supervisor information 231 can include can contain structured or unstructured supervisor information. For example, supervisor information can include a self-assessment of empathy on a 4-point scale, or a review by a patient that “this coach is very understanding.”

[0074] As a first nonlimiting example, supervisor information 231 may include information for a first supervisor:Name: SteveAge: 37Location: LondonPersonality: motivationalStyle: Actively encourage patients to improve.

[0075] As a second nonlimiting example, supervisor information 231 may include information for a second supervisor:Name: JenneyAge: 54Location: WisconsinPersonality: EmpatheticStyle: Seeks to form emotional connections with patients.

[0076] Supervisor interaction data 233 can include a record of the interactive opportunities selected by the supervisor and free-form or selectable text options for communications between the supervisor and a patient. For example, supervisor interaction data 233 may include an indication of how often a supervisor prefers to communicate with a patient, supervisor responsiveness (e.g., responding to a patient message within a certain amount of time or following up with another message if a period of time elapses between patient responses), or enthusiasm levels a supervisor would prefer to display during communication with a patient (e.g., the amount of exclamation points or emojis used). Supervisor interaction data 233 may also include a supervisor’s ranking of preference of types of messages (e.g., message templates).

[0077] FIG. 3 depicts exemplary contents of message bank 211, in accordance with disclosed embodiments. Message bank 211 can include message template 301. Such message templates can be provided by a supervisor and / or expert. Such message templates may be applicable to multiple patients, or multiple situations. In some embodiments, as shown, message template301 can include a message body 305 and message param eter(s) 303. In some instances, message template 301 can include only message body 305.

[0078] In some embodiments, a selection model can select a message template based on the message parameters associated with the message template. For example, the selection model can determine that a patient satisfied a rule that depends on a week-over-week change in activity level. In response to the satisfaction of that rule, the selection model can select a message template associated with that rule, or message templates including “week-over-week change in activity level” as a message parameter.

[0079] In some embodiments, message body 305 can include fixed content, such as a string of words or characters. In some embodiments, message parameter(s) 303 can specify variable content, such as the name of a patient or supervisor, a value relevant to the message, or the like. The variable content can be numeric, text, pictograms (e.g., emoji), or the like. The values of the variable content can be specific to a user, a supervisor, or a situation in a communication between a supervisor and a patient. The values of the variable content can be determined based on patient profiles 220 and / or supervisor profiles 230. For example, the values of the variable content specified by a message parameter can include information accessed from patient information 221, received from patient interaction data 223, accessed from supervisor information 231, received from supervisor interaction data 233, provided from an expert (e.g., via expert system(s) 106), or the like.

[0080] As may be appreciated, the disclosed embodiments are not limited to any particular format for specifying the mapping of message parameters to particular locations in the message body. For example, a message template associated with a detected drop in stress levels could include the following message body:{Great to see that your stress levels are now lower than last week. Hopefully, it will continue like that. }The message template could be selected, for example, based on a detected drop in stress levels recorded in a patient profile. The message template does not include any variable content and does not include any message parameters.

[0081] As an alternative example, another message template associated with a detected drop in stress levels could include the following message body:{Your stress levels decreased from %s last week to %s this week. Hopefully, it will continue like that. }As with the previous message, this message could be selected based on a detected drop in stress levels recorded in a patient profile. However, this message template does include variable content: the stress levels for each week. This message template could therefore also include the following message parameters:{ stress_level_prior_week, stress_level_current_week}When generating a parameterized message from the message template, application system 101 can determine values for the message parameters using the patient profile, and then insert the correct values into the message body.

[0082] In some embodiments, message param eter(s) 303 can be included in message body 305. For example:{I noticed you registered {num steps last week} steps last week. That's great to see! }In this example, the message parameter {num steps last week} is included in the message body. When generating a parameterized message from the message template, the correct value for num steps last week can replace the placeholder in the message body to generate the parameterized message.

[0083] Additional examples of message templates include:{{answers to survey}, You selected a couple of things that you'd like to gain from this program, such as: %s.}{{supervisor_name}, I'm here if you have any questions, %s! }

[0084] As may be appreciated, message parameters can be typed (e.g., typed as numeric, strings, images, or the like). In some embodiments, types can include objects like “supervisor name” or the like. Such types can be used in parameterizing a message template, to avoid placing variable content in an incorrect location in the message body.

[0085] In some embodiments, message bank 211 can include confidence values (e.g., confidence value 307) associated with stored message templates (e.g., message template 301). Such confidence values can indicate a confidence in the suitability, effectiveness, or the like of messages generated using message template 301. In some embodiments, application system 101 can be configured to preferentially select message templates based on confidence values. For example, a confidence value associated with a message can be a criterion considered by a selection model when selecting messages. In some embodiments, application system 101 can be configured to automatically send messages generated using message templates that satisfy a first confidence value condition (e.g., a high-confidence condition) without requiring supervisor approval. Otherwise, the messages may require supervisor approval to send to a patient. In some embodiments, message templates having a confidence value that satisfies a second confidence value condition (e.g., a low-confidence condition) can be selected or marked for review (e.g., by an expert or an administrator of application system 101). Such message templates may be revised or removed from message bank 211. For example, experts and / or supervisors may define an upper and a lower confidence value threshold. Messages generated using message templates associated with confidence values above the upper threshold (e.g., above 80 on a 0-100 scale or at a “high” or “very high” level on an ordinal scale) may be automatically sent (or automatically sent with a certainprobability). Message templates associated with confidence values below the lower threshold (e.g., below 20 on a 0-100 scale or at a “low” or “very low” level on an ordinal scale) may be automatically removed from the message bank, disabled from selection, automatically flagged for review, or some combination of the forgoing.

[0086] In some embodiments, confidence value 307 may be determined at least in part manually. For example, confidence value 307 can be set or adjusted based on supervisor and / or expert feedback. For example, an expert can set a confidence value for a message template (e.g., upon creation of the template, or subsequently). A supervisor can adjust the confidence value based on the supervisor’s assessment of the suitability, effectiveness, or the like of messages generated using the message template.

[0087] In some embodiments, confidence value 307 may be determined at least in part automatically. For example, confidence value 307 can be set or adjusted based on supervisorpatient interactions. Application system 101 can be configured to learn supervisors’ preferences for message templates (or messages generated from such message templates) and increase the confidence values for such messages. Application system 101 can learn such preferences based on the selection of messages from among sets of candidate messages, explicit ranking of messages, edits to messages, rejection of messages, or other supervisor interactions with messages.

[0088] The disclosed embodiments are not limited to any particular implementation of confidence value 307. Confidence value 307 may include continuous interval scales (e.g., ratings for integers on 0-100), ordinal scales (e.g., “very low confidence,” “low confidence,” “neutral confidence”, “high confidence,” “very high confidence”), or the like.

[0089] In some embodiments, message bank 211 can include selection rules (e.g., selection rules 309) associated with stored message templates (e.g., message template 301). Selection rules for a message template can evaluate to a Boolean value (e.g., using logical operations, acombination of numerical operations and express or implicit thresholding, or any combination of the foregoing) or a probability value based on information from a patient or supervisor profile. In some embodiments, a message template can be associated with a selection criterion (e.g., a set of if-then rules, Boolean expressions, or the like). Should information in the patient profile satisfy the selection criterion, then the corresponding message template may be considered for use in generating a message to the patient (or considered for use with a certain probability based on the probability value), as described herein. For example, a selection rule for a message template that congratulates a patient on a decrease in stress may be satisfied when the patient reports a decrease in stress. Similarly, a selection rule for a message template concerning portion size may be satisfied when a medical record of the patient indicates an increase in patient weight. Likewise, a selection rule for a message template that congratulates a patient on an increased activity level may be satisfied when data obtained from a patient device indicates that the patient took sufficient steps over a predetermined period. As may be appreciated, selection criterion can be or include inclusion criteria or exclusion criteria.

[0090] In some embodiments, message bank 211 can include category association(s) (e.g., category association(s) 311) associated with stored message templates (e.g., message template 301). A category association can indicate a strength of an association between the message template and a category (e.g., concerning a patient condition, competency, interaction, conversation part, or the like). In some embodiments, category association can be a tag or scalar associating the message template with a single category (e.g., sleep, diet, or the like). In some embodiments, the category association can be a list or vector indicating categories to which the message template belongs. In some embodiments, the values of the list or vector can be numeric, indicating a degree of the association between a message template and a category. For example, a message template may be highly associated with onepatient competency (e.g., sleep) but not highly associated with another patient competency (e.g., diet or exercise). The message template may have a greater numeric value for the highly associated competency than for the not-highly-associated competency.

[0091] In some embodiments, the category association(s) can be determined using, at least in part, manual labeling (e.g., by an expert that creates or updates the message template, or another user). In some embodiments, the category association(s) can be determined using, at least in part, a machine-learning model. For example, an embedding model can be trained to generate sentence embeddings that map message templates to a location in a space. Based on the location in the space, the message template can be associated with one or more categories. For example, a category can correspond to a dimension of the space or a location in the space. A projection of the message template location into the category dimension or a distance to the category location can determine the association between the message template and the category.

[0092] While confidence value 307, selection rules 309, and category association(s) 311 are described above as being stored in message bank 211, the disclosed embodiments are not so limited. In some embodiments, one or more of these values may be stored (e.g., in a database, table, repository, or the like) separately from message bank 211 (e.g., logically separately, or physically and logically separately on another system or device). In some embodiments, category association(s) 311 may not be stored. Instead, instructions or data for generating category associations for a message template can be stored in association with the message template. The instructions or data can be processed on-demand to generate the category associations. In some embodiments, the contents of message template 301 can constitute at least a part of such instructions or data. For example, the text of message template 301 can be used with an embeddings model to generate a sentence embedding on-demand.

[0093] FIG. 4 depicts a schematic of an exemplary system 400 for generating messages, consistent with disclosed embodiments. A patient-support platform can include or interact with system 400 to provide feedback messages to patients. For convenience of description, system 400 is depicted as being implemented using, or interacting with, components of architecture 100. For example, data stored in data storage 102 can be used to generate the messages. As an additional example, message generation component 401 can be implemented using, at least in part, application system 101. Similarly, message restatement component 403 can be implemented using, at least in part, machine learning system(s) 108. Message generation component 401 and message restatement component 403 can interact with data storage 102 to access recommendation model(s) 213, message bank 211, patient profiles 2203, supervisor profiles 230, or the like. Furthermore, user device 405 can be one of patient system(s) 103 or one of patient device(s) 104. However, system 400 can also be implemented using other databases, devices, or systems. The components of system 400 can be implemented using software, hardware, or any suitable combination of software and hardware.

[0094] Consistent with disclosed embodiments, message generation component 401 can be configured to generate messages for sending to patients. Message generation component 401 can be configured to generate messages using message templates obtained from message bank 211. In some embodiments, message generation component 401 can be configured to recommend message templates from message bank 211 using recommendation model(s) 213, patient profiles 220, and / or supervisor profiles 230.

[0095] In some embodiments, as described herein with regards to FIG. 7, message generation component 401 can perform a multi-phase recommendation process. In a first phase, candidate message templates can be identified using patient profile information and selection rules associated with the message templates (e.g., selection rules 309, or the like). In a secondphase, candidate message templates can be ranked or filtered using recommendation model(s) 213, patient profiles 220, and / or supervisor profiles 230. As may be appreciated, inclusion an initial rules-based selection phase can improve the controllability and interpretability of message generation component 401. Furthermore, recommendation model(s) 213 may be more computationally intensive than selection rules 309. Thus, limiting the set of candidate message templates using selection rules 309 prior to applying recommendation model(s) 213 may reduce the overall computational requirements of message generation component 401.

[0096] However, message generation component 401 is not limited to embodiments including such a multi-phase recommendation process. In some embodiments, message generation component 401 can perform a single-phase recommendation process using only recommendation model(s) 213 (or only selection rules 309). As may be appreciated, embodiments using only recommendation model(s) 213 may trade increased computational requirements and decreased controllability and interpretability against improved flexibility and accuracy. In contrast, embodiments using only selection rules 309 may trade decreased flexibility against reduced computational requirements and improved maintainability and scalability.

[0097] In some embodiments, message generation component 401 can be configured to use detection models to identify patients that require feedback. In some embodiments, the detection models can also identify suitable message templates. For example, a detection model can be trained to identify patients likely to qualify for a diagnosis of depression based on sleep, mood, and activity data. The detection model can trigger generation of a feedback message. In some embodiments, the detection model can also act as a recommendation model and suggest particular message templates for use in generating the feedback message to the patient (e.g., a message template concerning a mental health screening, or a message template concerning changes to diet, sleep, or physical activity).

[0098] In some embodiments, message generation component 401 can be configured to obtain feedback on recommended message templates. Message generation component 401 can receive such feedback from a supervisor (e.g., through one of supervisor system(s) 105). For example, message generation component 401 can provide multiple message templates for display to the supervisor (e.g., in a web browser or application). The supervisor can select among the message templates or re-rank the message templates, thereby providing feedback on the recommendation of message templates.

[0099] In some embodiments, message generation component 401 can use such feedback to update recommendation model(s) 213. Selection of a message template can increase the likelihood of future recommendation of the message template. Likewise, a failure to select a message template can decrease the likelihood of future recommendation of the message template. In this manner, message generation component 401 can implement supervised learning: training recommendation model(s) to recommend appropriate message templates.

[0100] In some embodiments, message generation component 401 can use such feedback to update confidence values associated with message templates. For example, a confidence value may be increased when a message template is displayed and selected, or decreased when the message template is displayed but not selected. In some such embodiments

[0101] In some embodiments, when a confidence value of a message template satisfies a confidence value condition (e.g., a high-confidence condition), system 400 may automatically provide a parameterized message generated using the message template to message restatement component 403. In some such embodiments, system 400 may automatically provide the parameterized message to message restatement component 403 with a probability p and present the message template or parameterized message for review with a probability 1- p, thus retaining some human oversight over the generated messages.

[0102] The disclosed embodiments are not limited to obtaining feedback on message templates. As may be appreciated, components of system 400 (e.g., message generation component 401, message restatement component 403, or the like) can additionally or alternatively obtain feedback on parameterized or restated messages. For example, a supervisor can select among or re-rank parameterized or restated messages, thereby generating training data for training recommendation model(s) 213 used to recommend message templates. Selection of a parameterized or restated message can increase the likelihood of recommendation of the message template used to generate the parameterized or restated message. Likewise, a failure to select a parameterized or restated message can decrease the likelihood of recommendation of the message template used to generate the parameterized or restated message.

[0103] In some embodiments, multiple versions of each message template may be present in message bank 211. In some embodiments, different version may be in different languages. The version recommended for a patient can depend on an express or estimated language preference of the patient. In some embodiments, when a supervisor and a patient speak different languages, the supervisor may review versions of message templates in the language of the supervisor while message generation component 401 generates parameterized messages using corresponding versions of the selected message templates in the language of the patient.

[0104] In some embodiments, different versions may correspond to different combinations of characteristics of the patient and the supervisor, such as relative status (e.g., honorific verb forms and vocabulary in Japanese), sex (e.g., as in German and similar languages), familiarity, or the like. The message generation component 401 can select a version of the message template based on these characteristics, which can be obtained from the patient profile 220 and / or the supervisor profile 230.

[0105] Consistent with disclosed embodiments, message templates can include a message body and may include message parameters. When generating a message for a patient using a message template, message generation component 401 can generate a parameterized message by inserting appropriate values for any such message parameters at the appropriate location in the message body. For example, given the message template:{{num_steps_last_week} {I noticed you registered %s steps last week. That's great to see! } and interaction data for a patient (e.g., stored in patient interaction data 223) that indicates that the patient has completed 12,500 steps last week (e.g., the variable num steps last week = 12500 for that patient), message generation component 401 can generate the parameterized message:I noticed you registered 12500 steps last week. That's great to see!Similarly, given the message template:{{supervisor_name}, I'm here if you have any questions, %s! } and supervisor information for a supervisor responsible for sending the message (e.g., stored in supervisor information 231) that indicates that the supervisor is named Jenney, (e.g., the variable supervisor name = “Jenney” for the supervisor responsible for sending the message), message generation component 401 can generate the parameterized message:I'm here if you have any questions, Jenney!

[0106] Consistent with disclosed embodiments, message restatement component 403 can be configured to restate messages for sending to patients. Message restatement component 403 can be configured to restate messages obtained from message generation component 401. In some instances, such a message can be a parameterized message. In some instances, such a message may include only a message body. In some embodiments, message restatementcomponent 403 can be configured to restate a message based on patient profile information and supervisor profile information.

[0107] In some embodiments, message restatement component 403 can restate a message by inputting the message to a machine-learning model. The machine-learning model can be configured (e.g., by training, prompting, or the like) to output a restated version of the message. The restated message may be semantically equivalent to the input message.However, the restated message may differ in language, presentation, or other characteristics from the input message. For example, the input message may be in an input language, while the restated language may be in an output language. Similarly, the input message may reflect a default combination of patient and supervisor characteristics (e.g., relative status, sex, etc.), while the actual combination of patient and supervisor characteristics differs from the default combination.

[0108] In some embodiments, configuring the machine-learning model can include generating a prompt that combines instructions and context information. The context information can be obtained from the patient profile and / or the supervisor profile. For example, system 400 can retrieve the patient language preference ’’English” and the supervisor sentiment preference “motivational and encouraging improvement” and construct a prompt as follows:“Re-state the following sentence(s) in English to be motivational and encouraging improvement:”This prompt includes instructions (e.g., “Re-state the following sentence(s)”) and context information obtained from the patient profile (e.g., “English”) and supervisory profile (e.g., “motivational and encouraging improvement”). The context information is combined with the instructions to generate the prompt (in this example by inserting the values of “English” and “motivational and encouraging improvement” into the prompt at predetermined locations).The prompt can then be concatenated with each parameterized message (or set of parameterized messages) to generate the following input to the machine-learning model. For example, the parameterized messages can be:1. I noticed you registered 12,500 steps last week. Good job.2. Great to see that your stress levels are now lower than last week. Hopefully, it will continue like that :)Then the input to the machine-learning model can be:“Re-state the following sentence(s) in English to be motivational and encouraging improvement:1. I noticed you registered 12,500 steps last week. Good job.2. Great to see that your stress levels are now lower than last week. Hopefully, it will continue like that :)”The resulting output of the machine-learning model can be, for example:1. Great job with 12,500 steps last week! That's impressive!2. I also noticed your stress levels are down. You're making great progress!In this manner, the output of the machine-learning model can preserve the content of the message while conforming to the supervisor-specified sentiment.

[0109] In some embodiments, message restatement component 403 can be configured to generate an output message in a particular output language using information retrieved from a patient profile (e.g., one of patient profiles 220). As shown above, in some instances a patient profile may expressly state a language. The message restatement component 403 may then obtain the patient language from the patient profile. In some instances, the message restatement component 403 may estimate a language based on a location of the patient (either expressly stated in the patient profile or obtained from a patient system or wearable device associated with the patient).

[0110] In some embodiments, message restatement component 403 can be configured to generate multiple output messages, each being a version of the input message in a different output language. A patient version of the output message can be in an express or estimated language of the patient. A supervisor version of the output message can be in an express or estimated language of the supervisor. In such embodiments, the supervisor can review and accept or reject the supervisor version, causing system 400 to send or discard the patient version. For example, the prompt can state: “Re-state the following sentence(s) from a supervisor to a patient in German to be motivational and encouraging improvement. Also provide an English translation of the restated sentence:”

[0111] In some embodiments, message restatement component 403 can be configured to generate an output message adapted to a particular combinations of characteristics of the patient and the supervisor, such as relative status, sex, familiarity, or the like. For example, a default configuration of the message template can assume a female patient and a female supervisor. The message restatement component 403 can determine that the grammar of the output language depends on the combination of the sexes of the patient and supervisor (e.g., German). The message restatement component 403 can determine that the patient and supervisor are male. The message restatement component 403 can be configured to generate an output message adapted to a male patient and a male supervisor. For example, a prompt can state: “Re-state the following sentence(s) from a supervisor to a patient in German to be motivational and encouraging improvement. The supervisor and the patient are male:” This prompt can be combined with the sentences to yield the input to the machine-learning model, as described herein.

[0112] In some embodiments, as described herein, the supervisor may have the opportunity to review a restated message (or a set of restated messages that form a communication). In some embodiments, the supervisor may accept, reject, or edit the restated messages. Theseinteractions can be used to fine-tune the machine-learning model used to restate the messages. For example, a training dataset can be created by tracking the choices presented to the supervisor and the interactions made by the supervisor. Consistent with disclosed embodiments, this training dataset can be used with Reinforcement Learning from Human Feedback (RLHF) or Direct Preference Optimization (DPO) training methods to bias the machine-learning model towards generated more preferred restated messages.

[0113] In some embodiments, when a confidence value of a message template used to generate the restated messages satisfies a first confidence value condition (e.g., a high- confidence condition), system 400 may automatically provide the restated messages (or communication) to the patient. In some such embodiments, system 400 may automatically provide the restated messages to the patient with a probability p and present the restated messages for review with a probability 1-p, thus retaining some human oversight over the generated messages.

[0114] Consistent with disclosed embodiments, user device 405 can be configured to receive restated messages from message restatement component 403. User device 405 can be a patient system (e.g., one of patient system(s) 103) or a patient device (e.g., one of patient device(s) 104) User device 405 can be configured to display the received restated message to the patient.

[0115] In some embodiments, measures of user engagement with the restated messages can be tracked. Such measures of user engagement can include whether the user opened the message, how long the user kept the message open, whether the user deleted the message, whether the user responded to the message, or the like. In some embodiments, system 400 can obtain such measures of user engagement (e.g., through a web browser or application executing on user device 405). Such measures of user engagement can provide another source of training data for use in fine-tuning the machine-learning model (or in selected messagetemplates) as described herein. For example, RLHF or DPO training methods can be used to bias the machine-learning model towards restating messages in a manner that encourages user engagement. Similarly, selection models can be updated to favor selection of message templates that result in higher levels of user engagement.

[0116] FIG. 5 A depicts an exemplary communication containing a set of message templates. Consistent with disclosed embodiments, such message templates can be obtained by message generation component 401, or a similar component. These message templates include message bodies showing insertion locations for the message parameters {num_steps_last_week}, {answers_to_survey} and {coach name}.

[0117] FIG. 5B to 5E depict exemplary restatements of parameterized messages generated from the message templates depicted in FIG. 5 A. These message templates demonstrate that the content of the message (e.g., as determined by the message templates) can be preserved, the sentiment of the message can be personalized to the supervisor, and message details can vary between restatements of the same message template. In this manner, the disclosed embodiments and support controlled, scalable generation of messages.

[0118] FIGs. 5B and 5C concern messages from the supervisor Steve to the patient Jane. In this example, Jane registered 12500 steps last week and wants to improve her sleep and lose weight. Message generation component 401 can generate parameterized messages using these parameter values and the message templates depicted in FIG. 5A.

[0119] FIGs. 5B and 5C depict two restatements of the parameterized messages generated using the parameter values for Jane and the message templates depicted in FIG. 5A. In these examples, Steve’s motivational and encouraging persona was used to configure message restatement component 403. The restated messages therefore display sentiment (e.g., mood, affect, emotion, valence, etc.) consistent with this persona.

[0120] In some embodiments, as depicted in FIGs. 5B and 5C, restatement need not be deterministic. Thus, two restatements of the same parameterized message templates can result in two different restated messages. As shown, both restated messages display the desired content and sentiment. However, they differ in phrasing and word choice. Such differences can improve the effectiveness of the restated messages, as patients are less likely to recognize or become accustomed to the underlying message template.

[0121] FIGs. 5D and 5E concern messages from the supervisor Jenney to the patient John. In this example, John registered 2000 steps last week and wants to recover from surgery and spend more time with his grandkids. Message generation component 401 can generate parameterized messages using these parameter values and the message templates depicted in FIG. 5 A.

[0122] FIGs. 5D and 5E depict two restatements of the parameterized messages generated using the parameter values for John and the message templates depicted in FIG. 5A. As given above, the supervisor Jenney is empathetic and seeks to form emotional connections with patients. Her emphatic persona was used to configure message restatement component 403. The restated messages therefore display sentiment (e.g., mood, affect, emotion, valence, etc.) consistent with this persona. As with the messages from Steve to Jane, the messages from Jenney to John retain the same content and sentiment but differ in phrasing and word choice.

[0123] FIG. 6 A depicts an exemplary communication containing a second set of message templates. Consistent with disclosed embodiments, such message templates can be obtained by message generation component 401, or a similar component. These message templates include message bodies showing insertion locations for the message parameters {num_fruits_this_week}, {num_vegs_this_week}, and {coach name}.

[0124] FIG. 6B and 6C depict exemplary restatements of parameterized messages generated from the message templates depicted in FIG. 6A. In these examples, Jane logged 7 fruit and18 vegetable servings, while John logged 3 fruit and 8 vegetable servings. FIG. 6B depicts a message from Steve to Jane, while FIG. 6C depicts a message from Jenney to John. The sentiments of these messages are consistent with the personas of the two supervisors, while the content reflects the message templates and the message parameter values.

[0125] In this manner, the disclosed system can achieve controlled, scalable distribution of messages. The content of a restated message can be controlled through selection of message templates (which can in turn be generated by experts), sentiment can be controlled by the supervisor, and presentation can be differ between instances of the restated message.

[0126] FIG. 7 depicts an exemplary method 700 for selecting message templates, consistent with disclosed embodiments. A patient-support platform can use method 700 to select message templates for use in generating feedback messages for patients. In some embodiments, components of system 400 can perform method 700. For example, for convenience of description, method 700 is described herein as being performed using a message generation component (e.g., message generation component 401, or the like) implemented using, or interacting with, components of architecture 100 (e.g., application system 101, data storage 102, supervisor system(s) 105, or the like). However, the disclosed embodiments are not so limited, as method 700 can also be performed using other systems or architectures.

[0127] In some embodiments, method 700 can be performed using message templates. The message templates can be obtained from a message bank (e.g., message bank 211, or the like). However, the disclosed embodiments are not so limited. In some embodiments, method 700 can be performed using message templates obtained from other sources. Additionally, in some embodiments, one or more steps of method 700 can be performed on parameterized messages or restated messages, as described herein. For example, message recommendation, filtering, or review can be performed on parameterized or restated messages.

[0128] In step 701 of method 700, the message generation component can select message templates, consistent with disclosed embodiments. The message generation component can select the message templates using information contained in a patient profile (e.g., patient profile 220, or the like). The information can include information about the patient (e.g., patient information 221, or the like) or interaction data for the patient (e.g., patient interaction data 223, or the like). In some embodiments, as described herein, message templates can be associated with selection rules. Such selection rules can evaluate to Boolean values or probabilities based on the contents of a patient profile. The message generation component can select the message templates using the associated selection rules.

[0129] In step 703 of method 700, the message generation component can recommend message templates, consistent with disclosed embodiments. In some embodiments, message generation component can recommend selected message templates using recommendation models (e.g., recommendation model(s) 213, or the like). As described herein, recommendation models can include if-then, statistical, or machine-learning models. The recommendation models can use patient profile information, category associations, sentence encodings, or the like to determine rankings of selected message templates. In some embodiments, the recommendation models can implement collective filtering using patient profiles and message templates. In some embodiments, the recommendation models can implement two-tower encoding models that map feature vectors for patients (e.g., based on patient profiles) and feature vectors for selected message templates to the same encoding space.

[0130] In some embodiments, the message generation component can determine a ranking of the selected message templates using the recommendation models. A number of the highest ranked message templates can then be recommended. In some embodiments, the message generation component can determine selection probabilities of the selected message templatesusing the recommendation models. A number of the message templates can then be chosen according to the selection probabilities.

[0131] In optional step 705 of method 700, the message generation component can filter messages. As may be appreciated, the message bank may include many similar messages. For example, different experts may unknowingly add similar messages. The message generation component can be configured to perform semantic filtering on the message templates selected in step 703. In some embodiments, category associations, sentence encodings, or the like can be used to assess message similarity. In this manner, the message generation component can obtain a set of dissimilar messages that are recommended for a patient and that satisfy selection rules associated with the messages.

[0132] While optional step 705 is depicted as occurring after step 703, the disclosed embodiments are not so limited. In some embodiments, optional step 705 can be performed before recommending messages in step 703. Such embodiments may benefit from a reduction in the computational resources required to perform step 703, as the filtering performed in step 705 may reduce the overall number of selected message templates.

[0133] In optional step 707 of method 700, the message generation component can display parameterized messages to a supervisor in a graphical user interface, such as the user interfaces displayed in FIGs. 7 to 9. As described herein, the supervisor can select or reorder the parameterized messages. The supervisor feedback can be used to update the recommendation models used in step 703, thereby improving the performance of method 700. As may be appreciated, the particular manner in which the recommendation models are updated can depend on the configuration of the recommendation model.

[0134] In some embodiments, when a confidence value of a message template used to generate a restated message satisfies a first confidence value condition (e.g., a high- confidence condition), the restated message can be automatically sent to the patient (e.g., step707 may be omitted). In some such embodiments, the restated message may be provided to the patient with a probability p and presented for review with a probability 1-p, thus retaining some human oversight over the generated messages.

[0135] FIG. 8 depicts an exemplary method 800 for generating messages and providing the messages to patients, consistent with disclosed embodiments. A patient-support platform can use method 800 to provide feedback messages to patients. In some embodiments, components of system 400 can perform method 800. For example, for convenience of description, method 800 is described herein as being performed using a message generation component (e.g., message generation component 401, or the like) and a method restatement component (e.g., message restatement component 403, or the like). The message generation component and message restatement component can be implemented using, or interacting with, components of architecture 100 (e.g., application system 101, data storage 102, supervisor system(s) 105, or the like). The generated messages can be provided to a user device (e.g., user device 405), which may be one of patient system(s) 103 or patient device(s) 104. However, the disclosed embodiments are not so limited, as method 800 can also be performed using other systems or architectures.

[0136] In step 801, method 800 can begin. In some embodiments, method 800 can begin according to a schedule. For example, method 800 can be performed periodically (e.g., every day hour, every day, every week, every month, or the like). In some embodiments, method 800 can be performed in response to an event. The event can be associated with a patient. For example, method 800 can be performed in response to a patient interaction or a change in patient information (e.g., as recorded in a patient profile 220, or the like). As may be appreciated, in some embodiments suitable detection models can be used to identify events requiring feedback based on patient interaction and / or a change in patient information.

[0137] Consistent with disclosed embodiments, patient interactions that trigger performance of method 800 can include, without limitation, acts or omissions such as: logging a meal in a meal journal by uploading text and a photo (or failing to log such a meal), completing (or failing to complete) an exercise mission, recording (or failing to record) adherence to a medication dosing schedule, attending (or failing to attend) a doctor’s appointment, requesting feedback from a supervisor regarding a program in which the user is enrolled, providing response to a questionnaire that indicate a change in the condition of the patient, or the like. For example, a patient-support platform could receive an upload from a patient of image of a meal. The patient-support platform could detect (e.g., using a multi-modal, machine-learning model) that the patient has eaten a healthy / unhealthy meal. In response, the patient-support platform could start method 800 to provide an appropriate feedback message to the patient.

[0138] Consistent with disclosed embodiments, changes in patient information that trigger performance of method 800 can include, without limitation, changes in medical records of the patient that indicate a change in a condition of the patient (e.g., indications of a flare-up of rheumatoid arthritis, gout, depression, or the like), patient monitoring data indicating sleep quality received from a patient device; records of alert signals received from an implanted pacemaker / defibrillator through an external controller; or the like. For example, patient monitoring data could indicate poor sleep quality and patient medical record data could indicate a diagnosis of depression. In response, the patient-support platform could start method 800 to provide a feedback message to the patient recommending the patient engage in additional exercise.

[0139] In step 810, system 400 can generate parameterized messages, consistent with disclosed embodiments. System 400 can generate the parameterized messages by selecting message templates from a message bank based on information contained in patient orsupervisor profiles. In some embodiments, the message templates can be selected using a multi-phase process, such as method 700 described herein with respect to FIGs. 4 and 7.

[0140] A first phase can be or include determining a set of candidate message templates, in some embodiments. The candidate message templates can be selected according to inclusion or exclusion criteria. In some embodiments, the inclusion or exclusion criteria for a message template may be associated with the message template. The information contained in the patient or supervisor profiles can be applied to the inclusion or exclusion criteria to determine whether to include the message template in the set of candidate message templates (or to determine a probability of including the message template in the set of candidate message templates).

[0141] A second phase can be or include determining ranking or recommending candidate message templates, in some embodiments. Recommendation models can be used to rank or recommend the candidate message templates. For example, a recommendation model can score candidate message templates. A number of the highest-scoring candidate message templates, or candidate message templates having scores exceeding a threshold, can be selected for use in generating messages to patients. As described herein, rule-based, statistical, or machine-learning models (or a combination thereof) can be used to rank or recommend the candidate message templates.

[0142] In some embodiments, recommended messages can be semantically filtered. Such semantical filtering can include grouping recommended messages that satisfy a similarity criterion. The similarity criterion can depend on category associations, sentence embeddings, or the like. For example, the similarity criterion can be a cosine similarity between sentence embeddings of pairs of recommended messages. As an additional example, the similarity criterion can be a nearest-neighbor grouping of embedding locations generated usingcategory associations. For each grouping, a representative recommended message can be selected for use in generating messages to patients.

[0143] While described with regards to recommended messages, the disclosed embodiments are not so limited. In some embodiments, semantic filtering can be performed on candidate messages.

[0144] Consistent with disclosed embodiments, the message templates can be parameterized using values obtained from the patient or supervisor profiles for (any) message parameters included in the message template. Such parameterization can occur before the first phase, after the first phase and before the second phase, or after the second phase. Similarly, such parameterization can occur before or after semantic filtering.

[0145] In step 820, parameterized messages can be restated. As described herein, restating the parameterized messages can improve engagement by maintaining novelty and avoiding a sense of repetition in the patient. A parameterized message can be restated, as described with respect to FIG. 4, using a machine-learning model. The machine-learning model can be configured (e.g., by training, prompting, or the like) to output a restated version of the message. The restated message may be semantically equivalent to the input message.However, the restated message may differ in language, presentation, or other characteristics from the input message. Thus, the content of the message can be preserved, the sentiment of the message can be personalized to the supervisor, and message details can vary between restatements of the same message template. In this manner, the disclosed embodiments and support controlled, scalable generation of messages.

[0146] In some embodiments, message can be restated into a language of the patient. In this manner, a system consistent with disclosed embodiments can be configured to use a single language for message templates and prompt generation, while supporting the provision of messages to patients (and supervisors) speaking a variety of languages.

[0147] In step 830, the restated messages can be provided to patients. The disclosed embodiments are not limited to any particular channel, format, or protocol for providing the restated messages. In some instances, a restated message can be provided to a patient system, while in other instances, the restated message can be provided to a patient device. The message can be provided for display in a user interface, such as a web browser or application executing on the patient system or patient device. In some embodiments, messages can be converted to audio, or can be used as prompts for multi-modal generative machine-learning models to generate images.

[0148] In some embodiments, method 800 can include steps of obtaining supervisor feedback. In various embodiments, supervisor feedback may be obtained in step 841 when generating parameterized messages or in step 843 when restating messages using a machinelearning model.

[0149] In step 841, supervisor feedback can include selecting among candidate message templates or parameterized messages, or rearranging candidate message templates or parameterized messages. The supervisor feedback can be used to update recommendation model(s) used to generate the candidate message templates, or the message templates used to generate the parameterized messages. In some embodiments, supervisor feedback regarding a message template or parameterized message can be omitted when a confidence value associated with the message template, or the message template used to generate the parameterized message, satisfies a confidence value condition (e.g., a high-confidence value condition, or the like).

[0150] In step 843, supervisor feedback can include selecting among restated messages, rearranging restated messages, or editing restated messages. In some embodiments, supervisor feedback regarding a restated message template can be omitted when a confidence value associated with the message template, or the message template used to generate theparameterized message, satisfies a confidence value condition (e.g., a high-confidence value condition, or the like). In some embodiments, the supervisor feedback can be used to update recommendation model(s) used to generate the message templates used to generate the restated messages. In some embodiments, the supervisor feedback can be used to train or fine-tune the machine-learning model used to restate the messages.

[0151] In some embodiments, supervisor feedback may be provided only in step 843. In such embodiments, multiple candidate parameterized messages can be generated in step 810. The candidate parameterized messages can then be restated in step 820. The candidate parameterized messages can then be reviewed in by a supervisor using a graphical user interface (e.g., as depicted in FIGS. 9, 10A, 10B, and 11). The supervisor can then select among the restated messages, rearrange the restated messages, or edit the restated messages. The review can then approve the selected, rearranged, or edited message(s). The message(s) can then be sent to the patient.

[0152] In step 899, method 800 can terminate. In some embodiments, system 400 can perform another iteration of method 800 for another patient. In some embodiments, system 400 can provide engagement feedback to supervisors that messages were sent, received, opened, etc. In some embodiments, system 400 can provide feedback to experts regarding message template confidence values (e.g., flagging messages that satisfy a low-confidence condition), supervisor selection or rearrangement message templates, parameterized messages, or restated messages, or the editing of restated messages.

[0153] FIG. 9 depicts an exemplary supervisor interface 900 for reviewing messages, consistent with disclosed embodiments. The disclosed embodiments are not limited to any particular implementation of supervisor interface 900. In some embodiments, supervisor interface 900 can be part of an application running on a supervisor system (e.g., one ofsupervisor system(s) 105). In some embodiments, supervisor interface 900 can be provided by application system 101 to the supervisor system through a web browser or the like.

[0154] Consistent with disclosed embodiments, supervisor interface 900 can enable a supervisor to select, modify, rank, or approve messages generated by application system 101. Supervisor interface 900 can include a recommended message field 902. Recommended message field 902 can display a candidate communication. The communication can include a set of messages to the patient. In this example, the communication includes restated versions of messages (e.g., restated versions of parameterized messages). The restated messages may be appropriately phrased, given the language, gender, relative status, or the like for the supervisor and patient in the communication. The restated messages may also convey a sentiment selected by the supervisor. The supervisor can approve the candidate communication, causing application system 101 to send the candidate communication (and the restated messages included therein) to the patient.

[0155] In some embodiments, application system 101 can be configured to generate multiple candidate communications. In such embodiments, supervisor interface 900 may include a navigation control 904. The navigation control can enable the supervisor to navigate between candidate communications. For example, supervisor interface 900 may recommend 1, 2, 3, or more communications, and a supervisor can interact with navigation control 904 to view these different candidate communications. The supervisor can then approve a candidate communication, causing application system 101 to send the approved candidate communication to the patient.

[0156] In some embodiments, supervisor interface 900 can include a modification control 906. The modification control can enable the supervisor to edit restated message included within a communication using recommended message field 902. In some embodiments,application system 101 can track the edits. As described herein, such feedback can be used to train or fine tune a machine-learning model used to restate the parameterized messages.

[0157] FIGs. lOA to 10B depict an exemplary supervisor interface 1000 for reviewing messages, in accordance with disclosed embodiments. The disclosed embodiments are not limited to any particular implementation of supervisor interface 1000. In some embodiments, supervisor interface 1000 can be part of an application running on a supervisor system (e.g., one of supervisor system(s) 105). In some embodiments, supervisor interface 1000 can be provided by application system 101 to the supervisor system through a web browser or the like.

[0158] Consistent with disclosed embodiments, supervisor interface 1000 may include one or more restated messages, as described herein. The restated messages may be formatted for agreement with proper grammar, language, gender, or the like for the supervisor and patient in the communication. Restated messages displayed on supervisor interface 1000 may be selected or ranked according to application system 101, as described herein.

[0159] In some embodiments, supervisor interface 1000 may include various categories of restated messages. As described herein, such categories can concern patient conditions, patient competencies, patient interactions, conversation parts, or the like. In some embodiment, restated messages can be associated with categories based on the message templates used the generate the restated messages. In the example displayed in FIGs. 10A and 10B, restated messages are organized by conversation parts and certain patient competencies. The categories displayed include greeting field 1002, task field 1004, move field 1006, food field 1008, question field 1010, and feedback field 1012. Greeting field 1002 may include restated messages corresponding to a greeting to be sent to a patient. Task field 1004 may include restated messages corresponding to a task or reminder the supervisor may request of the patient. Move field 1006 may include restated messages corresponding to an activity orexercise the supervisor may request of the patient. Food field 1008 may include restated messages related to food or diet for the patient. Question field 1010 may include restated messages related to a question the supervisor may ask of the patient. Feedback field 1012 may include restated messages relating to feedback the supervisor provides to the patient, including negative and / or positive feedback. Additionally, or alternatively, supervisor interface 1000 may include a miscellaneous field for other restated messages. As may be appreciated, supervisor interface 1000 can be configured to include fields corresponding to other categories, or other arrangements of categories.

[0160] In some embodiments, supervisor interface 1000 can be configured with controls (e.g., checkboxes or radio buttons, or the like) to enable a supervisor may select one or more messages from the various fields 1002-1012. In some embodiments, supervisor interface 1000 may include an auto-select field 1001. Auto-select field 1001 may automatically choose one or more messages from fields 1002-1012. For example, auto-select field 1001 may automatically select the highest-ranked message from fields 1002-1012 (e.g., within each category). The ranking can be determined using selection models, as described herein. In this example, selections can be cleared using a “clear selections” control.

[0161] FIG. 11 depicts an exemplary supervisor interface 1100 for arranging message templates, in accordance with disclosed embodiments. The disclosed embodiments are not limited to any particular implementation of supervisor interface 1100. In some embodiments, supervisor interface 1100 can be part of an application running on a supervisor system (e.g., one of supervisor system(s) 105). In some embodiments, supervisor interface 1100 can be provided by application system 101 to the supervisor system through a web browser or the like.

[0162] Consistent with disclosed embodiments, supervisor interface 1100 may allow supervisor interaction to sort message templates, parameterized messages, and / or restatedmessages. For example, supervisor interface 1100 may allow a supervisor to sort messages selected from message field 1002-1012, as referenced in FIGS. 10A and 10B. Supervisor interface 1100 may include a sorting field 1101. Sorting field 1101 may include selected restated messages from message fields 1002-1012, as well as other messages, such as a restated greeting message 1102, restated task message 1104, restated move message 1106, restated food message 1108, restated question message 1110, and closing message 1112. A supervisor may interact with sorting field 1101 to rearrange, add, delete, or modify the messages. For example, a supervisor may long-click and drag a message in sorting field 1101 as desired to rearrange an ordering of the messages.

[0163] As may be appreciated, once the supervisor has selected and sorted restated messages (e.g., using supervisor interfaces 1000 and 1100, or the like), an overall message can be displayed to the supervisor (e.g., using an interface similar to supervisor interface 900). The user may then review (and optionally edit) the message, prior to approving the message for sending by application system 101 to the patient. Edits to the message can be used to adjust confidence values of message templates used to generate the messages, consistent with disclosed embodiments.

[0164] As used herein, unless specifically stated otherwise, the term “or” encompasses all possible combinations, except where infeasible. For example, if it is stated that a component may include A or B, then, unless specifically stated otherwise or infeasible, the component may include A, or B, or A and B. As a second example, if it is stated that a component may include A, B, or C, then, unless specifically stated otherwise or infeasible, the component may include A, or B, or C, or A and B, or A and C, or B and C, or A and B and C.

[0165] Additionally, the articles “a” and “an” as used in the present disclosure and the appended claims should generally be construed to mean “one or more” unless specified otherwise or clear from context to be directed to a singular form.

[0166] The flowcharts and block diagrams in the figures illustrate examples of the architecture, functionality, and operation of possible implementations of systems, methods, and devices according to various embodiments. It should be noted that, in some alternative implementations, the functions noted in blocks may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. Likewise, additional steps may be included in such methods, and certain steps may be omitted or combined, in methods consistent with various embodiments.

[0167] It is understood that the described embodiments are not mutually exclusive, and elements, components, materials, or steps described in connection with one example embodiment may be combined with, or eliminated from, other embodiments in suitable ways to accomplish desired design objectives.

[0168] The disclosed embodiments may further be described using the following clauses:

[0169] 1. A message-generation system comprising: at least one processor; and at least one computer-readable medium containing instructions that, when executed by the at least one processor, cause the message-generation system to perform operations comprising: in accordance with a schedule or in response to an event associated with a user, generating a parameterized message using a message template; generating a restated message by applying the parameterized message to a machine-learning model configured to restate the parameterized message; and providing the restated message to the user.

[0170] 2. The message-generation system of claim 1, wherein: generating the parameterized message includes performing a multi-phase recommendation process using patient profile information: a first phase including selecting candidate message templates or parameterized messages using selection rules and patient profile information; and a second phase includingrecommending a selected candidate message template or parameterized message using a recommendation model.

[0171] 3. The message-generation system of any one of clauses 1 to 2, wherein: the operations further comprise: obtaining supervisor feedback concerning the recommended candidate message template or parameterized message; and updating the recommendation model based on the obtained supervisor feedback.

[0172] 4. The message-generation system of any one of clauses 1 to 3, wherein: generating the parameterized message includes recommending a candidate message template or parameterized message using: at least one category association associated with the candidate message template or parameterized message; or an embedding associated with the candidate message template or parameterized message.

[0173] 5. The message-generation system of any one of clauses 1 to 3, wherein: generating the parameterized message includes recommending a candidate message template or parameterized message using: collaborative filtering of patients and message templates or parameterized messages, or two-tower encoding model of the patients and the message templates or parameterized messages.

[0174] 6. The message-generation system of any one of clauses 1 to 5, wherein: the message template, parameterized message, or restated message is associated with a confidence value; and generating the parameterized message or restated message includes: determining that the parameterized message or restated message requires review based on the confidence value; and obtaining, in response to the determination, supervisor approval of the parameterized message or restated message.

[0175] 7. The message-generation system of any one of clauses 1 to 5, further comprising: the message template, parameterized message, or restated message is associated with a confidence value; and generating the parameterized message or restated message includesdetermining that the parameterized message or restated message satisfies an automatic sending criterion based on the confidence value:

[0176] 8. The message-generation system of any one of clauses 1 to 7, wherein: the machinelearning model is configured to generate the restated message by restating the parameterized message using a patient profile or a supervisory profile.

[0177] 9. The message-generation system of claim 8, wherein: the operations further comprise: obtaining supervisor feedback concerning the restated message; and updating or fine-tuning the machine-learning model using the obtained supervisor feedback.

[0178] 10. The message-generation system of any one of clauses 1 to 9, wherein: generating the restated message comprises translating the parameterized message based on patent profile or supervisory profile information.

[0179] 11. The message-generation system of any one of clauses 1 to 10, wherein: configuring the machine-learning model comprises: generating a prompt including: context information comprising at least some of a patent profile or at least some of a supervisor profile; and instructions to restate the parameterized message using the context information; and applying the prompt to the machine-learning model together with the parameterized message.

[0180] 12. The message-generation system of any one of clauses 1 to 11, wherein: the parameterized message is generated in response to: receiving an image from the user; and detecting, using a detection machine-learning model, one or more items in the image.

[0181] 13. The message-generation system of any one of clauses 1 to 12, wherein: generating the parameterized message includes semantic filtering of candidate message templates or parameterized messages.

[0182] 14. A message-generation system comprising: at least one processor; and at least one computer-readable medium containing instructions that, when executed by the at least oneprocessor, cause the message-generation system to perform operations comprising: generating a restated message using a machine-learning model, a recommendation model, a message template, a profile of a patient, and a supervisor profile; displaying, in a first view of a supervisor interface, a communication including the restated message or multiple selectable restated messages including the restated message; and in response to an interaction of a supervisor with the supervisor interface, providing the restated message to the patient.

[0183] 15. The message-generation system of claim 14, wherein: the first view displays the communication and a control selectable to generate another version of the communication.

[0184] 16. The message-generation system of claim 14, wherein: the first view displays the communication and a control selectable to: enable editing of the communication; and enable training or fine-tuning of the machine-learning model based on the editing of the communication.

[0185] 17. The message-generation system of claim 14, wherein: the first view displays the multiple selectable restated messages; and the operations further comprise: in response to selection of ones of the multiple selectable restated messages, displaying, in a second view of the supervisor interface, a re-orderable list of the selected ones of the multiple selectable restated messages.

[0186] 18. The message-generation system of claim 14, wherein: the first view displays the restated message selectable to: enable training of the recommendation model based on the selection of the restated message.

[0187] 19. The message-generation system of any one of clauses 14 to 18, wherein: generating the restated message comprises: generating a parameterized message; and restating the parameterized message using the machine-learning model, the patient profile, and the supervisor profile.

[0188] 20. The message-generation system of any one of clauses 14 to 19, wherein: generating the parameterized message comprises performing a multi-phase recommendation process, comprising: a first phase including selecting the message template or parameterized message using selection rules and the patient profile; and a second phase including recommending the selected message template or parameterized message using the recommendation model.

[0189] It will be further understood that various modifications, alternatives, and variations in the details, materials, and arrangements of the parts which have been described and illustrated to explain the nature of described embodiments may be made by those skilled in the art without departing from the scope. Accordingly, the following claims embrace all such alternatives, modifications, and variations that fall within the terms of the claims.

Claims

WHAT IS CLAIMED IS:

1. A message-generation system comprising: at least one processor; and at least one computer-readable medium containing instructions that, when executed by the at least one processor, cause the message-generation system to perform operations comprising: in accordance with a schedule or in response to an event associated with a user, generating a parameterized message using a message template; generating a restated message by applying the parameterized message to a machine-learning model configured to restate the parameterized message; and providing the restated message to the user.

2. The message-generation system of claim 1, wherein: generating the parameterized message includes performing a multi-phase recommendation process using patient profile information: a first phase including selecting candidate message templates or parameterized messages using selection rules and patient profile information; and a second phase including recommending a selected candidate message template or parameterized message using a recommendation model.

3. The message-generation system of claim 2, wherein: the operations further comprise: obtaining supervisor feedback concerning the recommended candidate message template or parameterized message; and updating the recommendation model based on the obtained supervisor feedback.

4. The message-generation system of claim 1, wherein: generating the parameterized message includes recommending a candidate message template or parameterized message using: at least one category association associated with the candidate message template or parameterized message; or an embedding associated with the candidate message template or parameterized message.

5. The message-generation system of claim 1, wherein: generating the parameterized message includes recommending a candidate message template or parameterized message using: collaborative filtering of patients and message templates or parameterized messages, ora two-tower encoding model of the patients and the message templates or parameterized messages.

6. The message-generation system of claim 1, wherein: the message template, parameterized message, or restated message is associated with a confidence value; and generating the parameterized message or restated message includes: determining that the parameterized message or restated message requires review based on the confidence value; and obtaining, in response to the determination, supervisor approval of the parameterized message or restated message.

7. The message-generation system of claim 1, wherein: the message template, parameterized message, or restated message is associated with a confidence value; and generating the parameterized message or restated message includes determining that the parameterized message or restated message satisfies an automatic sending criterion based on the confidence value:

8. The message-generation system of claim 1, wherein: the machine-learning model is configured to generate the restated message by restating the parameterized message using a patient profile or a supervisory profile.

9. The message-generation system of claim 8, wherein: the operations further comprise: obtaining supervisor feedback concerning the restated message; and updating or fine-tuning the machine-learning model using the obtained supervisor feedback.

10. The message-generation system of claim 1, wherein: generating the restated message comprises translating the parameterized message based on patent profile or supervisory profile information.

11. The message-generation system of claim 1, wherein: configuring the machine-learning model comprises: generating a prompt including: context information comprising at least some of a patent profile or at least some of a supervisor profile; and instructions to restate the parameterized message using the context information; and applying the prompt to the machine-learning model together with the parameterized message.

12. The message-generation system of claim 1, wherein: the parameterized message is generated in response to: receiving an image from the user; and detecting, using a detection machine-learning model, one or more items in the image.

13. The message-generation system of claim 1, wherein: generating the parameterized message includes semantic filtering of candidate message templates or parameterized messages.

14. A message-generation system comprising: at least one processor; and at least one computer-readable medium containing instructions that, when executed by the at least one processor, cause the message-generation system to perform operations comprising: generating a restated message using a machine-learning model, a recommendation model, a message template, a profile of a patient, and a supervisor profile; displaying, in a first view of a supervisor interface, a communication including the restated message or multiple selectable restated messages including the restated message; andin response to an interaction of a supervisor with the supervisor interface, providing the restated message to the patient.

15. The message-generation system of claim 14, wherein: the first view displays the communication and a control selectable to generate another version of the communication.

16. The message-generation system of claim 14, wherein: the first view displays the communication and a control selectable to: enable editing of the communication; and enable training or fine-tuning of the machine-learning model based on the editing of the communication.

17. The message-generation system of claim 14, wherein: the first view displays the multiple selectable restated messages; and the operations further comprise: in response to selection of ones of the multiple selectable restated messages, displaying, in a second view of the supervisor interface, a re-orderable list of the selected ones of the multiple selectable restated messages.

18. The message-generation system of claim 14, wherein:the first view displays the restated message selectable to: enable training of the recommendation model based on the selection of the restated message.

19. The message-generation system of claim 14, wherein: generating the restated message comprises: generating a parameterized message; and restating the parameterized message using the machine-learning model, the patient profile, and the supervisor profile.

20. The message-generation system of claim 19, wherein: generating the parameterized message comprises performing a multi-phase recommendation process, comprising: a first phase including selecting the message template or parameterized message using selection rules and the patient profile; and a second phase including recommending the selected message template or parameterized message using the recommendation model.

Images