Method and system for generating chronic disease intervention scheme based on reinforcement learning and multi-modal data

Abstract

The invention relates to the technical field of reinforcement learning, and discloses a chronic disease intervention scheme generation method and system based on reinforcement learning and multi-modal data, and the method comprises the steps: obtaining the multi-modal data of a patient, the multi-modal data at least comprising electronic medical record data, voice data, image data, text data and physiological time sequence data; performing feature extraction on the multi-modal data to obtain a multi-modal feature vector; on the basis of the multi-modal feature vectors, health state vectors are constructed, and the health state vectors at least comprise a physiological risk score, a treatment compliance score and a lifestyle health degree score; designing a reward function according to the dynamic change of the health state vector; optimizing the strategy network in a predefined intervention action space according to the reward function by utilizing a reinforcement learning algorithm so as to output an optimal intervention action; and converting the optimal intervention action into personalized natural language interaction content through a generative AI model. According to the invention, the efficiency, precision and patient compliance of chronic disease management can be significantly improved.

Description

Technical Field

[0001] This application relates to the field of reinforcement learning technology, and in particular to a method and system for generating chronic disease intervention programs based on reinforcement learning and multimodal data. Background Technology

[0002] Chronic non-communicable diseases (such as cardiovascular and cerebrovascular diseases and diabetes) have become a major public health problem in my country, accounting for up to 74% of all deaths. County-level primary healthcare institutions, as the main implementers of chronic disease management, face the challenges of heavy follow-up tasks and low management efficiency.

[0003] Currently, the manual follow-up model widely used at the grassroots level suffers from low efficiency and insufficient coverage. The collected data is mostly in unstructured form (such as dialectal oral accounts and handwritten records), making it difficult to analyze and utilize effectively. Although some AI-assisted tools attempt to improve this situation, they still have significant limitations: at the data level, they mostly use single-modal or simple multimodal data stacking, lacking the ability to deeply integrate and cross-validate heterogeneous data such as voice, images, and text; at the model level, they mainly rely on static models based on rules or supervised learning, unable to dynamically adjust intervention strategies according to real-time patient data, resulting in insufficient personalization; at the interaction level, existing systems have mechanical and rigid interaction methods, making it difficult to perceive the patient's emotional state, leading to low patient compliance. Summary of the Invention

[0004] The purpose of this invention is to address at least one deficiency in the existing technology and to provide a method and system for generating chronic disease intervention programs based on reinforcement learning and multimodal data. This invention can significantly improve the efficiency, accuracy and patient compliance of chronic disease management.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for generating chronic disease intervention programs based on reinforcement learning and multimodal data, the method comprising the following steps: Acquire multimodal data of patients, including at least electronic medical record data, voice data, image data, text data, and physiological time-series data from wearable devices; Feature extraction is performed on the multimodal data to obtain multimodal feature vectors; Based on the multimodal feature vector, a health status vector is constructed, which includes at least a physiological risk score, a treatment adherence score, and a lifestyle health score. Design a reward function based on the dynamic changes of the health state vector; Using a reinforcement learning algorithm, the policy network is optimized in a predefined intervention action space according to the reward function to output the optimal intervention action; The optimal intervention action is transformed into personalized natural language interactive content through a generative AI model.

[0006] Furthermore, the feature extraction of the multimodal data specifically includes: Using a pre-trained language model in the medical field, medical entities, sentiment tendencies, and user intent features are extracted from the speech data and the text data. The medical entities include one or more of symptoms, drugs, and physiological indicators. Acoustic features are extracted from the speech data, and based on the acoustic features, the patient's anxiety and fatigue emotional state are quantified using speech emotion recognition technology. The acoustic features include pitch, speech rate, and energy. Numerical information is extracted from the image data using OCR technology.

[0007] Furthermore, the physiological risk score is calculated using the following formula:

[0008] in, For physiological risk scoring, Let i be the clinical weight of the i-th physiological indicator. Let i be the current measured value of the i-th indicator. Let be the clinical control target value for the i-th physiological indicator. This is a normalization function used to calculate the degree of deviation between the current value and the target value; Furthermore, the treatment adherence score is calculated using the following formula:

[0009] in, For treatment compliance scores, Medication adherence is calculated based on the ratio of actual to required doses. Dietary adherence is calculated based on the match between patient meal photos and recommended nutritional plans. Exercise adherence is calculated based on the match between patient exercise data and exercise prescription. , , The weighting coefficients are dynamically adjusted based on the patient's condition, and satisfy the following conditions: = 1.

[0010] Furthermore, the reward function is calculated using the following formula:

[0011] in, For the reward function, , , These represent the positive changes in the ratings at adjacent time points, and = , = , = , Rate your lifestyle health. , , These are the weighting coefficients, and C is the interaction cost constant.

[0012] Furthermore, weighting coefficients , , The weighting is dynamically adjusted based on the patient's condition, which includes newly diagnosed patients, patients with unstable conditions, patients with poor compliance, or patients in a stable phase. For newly diagnosed patients, the weighting coefficient is adjusted accordingly. Greater than the weight coefficient and weighting coefficients For patients with unstable conditions, the weighting coefficient Greater than the weight coefficient and weighting coefficients For patients with poor compliance, the weighting coefficient Greater than the weight coefficient and weighting coefficients For patients in the stable phase, the weighting coefficient Greater than the weight coefficient and weighting coefficients .

[0013] Furthermore, after feature extraction of the multimodal data, a multimodal data cross-validation step is also included: comparing the same type of data from different modalities, and if the difference exceeds a preset threshold, a follow-up questioning procedure or manual verification process is initiated.

[0014] Furthermore, the health status vector also includes an emotion score obtained based on speech data recognition or text data analysis.

[0015] Secondly, the present invention also provides a chronic disease intervention program generation system based on reinforcement learning and multimodal data, the system being based on the method described in the first aspect, comprising: Acquisition module: configured to acquire multimodal data of patients, including at least electronic medical record data, voice data, image data, text data, and physiological time-series data from wearable devices; Feature extraction module: configured to extract features from the multimodal data to obtain multimodal feature vectors; The construction module is configured to construct a health status vector based on the multimodal feature vector, wherein the health status vector includes at least a physiological risk score, a treatment adherence score, and a lifestyle health score. Design module: configured to design a reward function based on the dynamic changes of the health state vector; Output module: configured to use a reinforcement learning algorithm to optimize a policy network in a predefined intervention action space based on the reward function, so as to output the optimal intervention action; The conversion module is configured to convert the optimal intervention into personalized natural language interactive content using a generative AI model.

[0016] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in the first aspect.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention integrates multimodal data and constructs a quantified health state vector. It uses a reward function centered on the improvement in state to drive the optimization of reinforcement learning strategies, enabling dynamic, personalized, and interpretable chronic disease intervention. This effectively solves the problems of single data, static recommendations, and lack of continuous optimization in traditional follow-up methods, significantly improving the efficiency, accuracy, and patient compliance of chronic disease management. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for generating chronic disease intervention programs based on reinforcement learning and multimodal data, according to Embodiment 1 of the present invention. Figure 2 Here is a system block diagram for generating a chronic disease intervention scheme based on reinforcement learning and multimodal data, as described in Embodiment 3 of this invention. Figure 3 This is a flowchart of the multimodal data fusion and feature extraction process in Embodiment 1 of the present invention; Figure 4 This is a flowchart of the intervention decision-making process based on reinforcement learning in Embodiment 1 of the present invention; Figure 5 This is a flowchart illustrating the generative AI-driven personalized solution generation and interaction process of Embodiment 1 of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] Example 1 Please see Figure 1 A preferred embodiment of the present invention provides a method for generating chronic disease intervention programs based on reinforcement learning and multimodal data, the method comprising the following steps: S1: Acquire the patient's multimodal data, which includes at least electronic medical record data, voice data, image data, text data, and physiological time-series data from wearable devices; S2: Perform feature extraction on the multimodal data to obtain a multimodal feature vector; S3: Based on the multimodal feature vector, construct a health status vector, which includes at least a physiological risk score, a treatment compliance score, and a lifestyle health score; S4: Design a reward function based on the dynamic changes of the health state vector; S5: Using a reinforcement learning algorithm, optimize the policy network in a predefined intervention action space according to the reward function to output the optimal intervention action; S6: Transform the optimal intervention action into personalized natural language interactive content using a generative AI model.

[0021] This embodiment integrates multimodal data and constructs a quantified health state vector. It uses a reward function with state improvement as the core to drive reinforcement learning strategy optimization, which enables dynamic, personalized and interpretable chronic disease intervention. It effectively solves the problems of single data, static suggestions and lack of continuous optimization in traditional follow-up methods, and significantly improves the efficiency, accuracy and patient compliance of chronic disease management.

[0022] In an optional embodiment, in step S1, electronic medical record (EHR) data can be obtained through cooperation with medical institutions such as hospitals and community health service centers, and with patient authorization, by retrieving it from a standardized interface. Optionally, in some regions where a regional health information platform has been established, which aggregates health data from multiple medical institutions within the region, the relevant EHR data can also be accessed through an authorized interface, provided that patient authorization is obtained and platform management regulations are met. In this embodiment, EHR data specifically includes structured or semi-structured data such as the patient's basic information, outpatient / inpatient records, diagnostic results, medical orders, medication records, laboratory test reports, and medical image indexes. The voice data, image data, and text data in the multimodal data are all taken from follow-up data.

[0023] This embodiment directly retrieves the EHR through a standardized interface, ensuring the accuracy and authority of historical medical data (such as diagnostic results and medication records) and avoiding errors from manual entry. Simultaneously, it combines dynamic data such as voice, images, and text obtained from follow-up with the static EHR to create a comprehensive and multi-dimensional description of the patient's health status, providing strong data support for the subsequent construction of a high-precision health status vector.

[0024] In an optional embodiment, please refer to Figure 3 In step S2, the feature extraction of the multiple modal data specifically includes: S2.1: Utilizing a pre-trained language model in the medical field, medical entities, sentiment tendencies, and user intent features are extracted from the speech and text data. The medical entities include one or more of symptoms, medications, and physiological indicators. In this embodiment, the pre-trained language model in the medical field is ClinicalBERT. Compared to general language models, ClinicalBERT can more accurately identify and understand medical terms, drug names, and complex symptom descriptions (such as 'paroxysmal vertigo'), significantly reducing the error rate of medical entity extraction and providing highly reliable structured information for subsequent quantitative evaluation. Furthermore, the 'medical entities' (such as symptoms and medications) extracted using professional models like ClinicalBERT are themselves concepts understandable in the medical field. This makes the system's internal decision-making process (such as adjusting intervention strategies based on the aggravation of specific symptoms) more interpretable for doctors, helping to increase the trust and acceptance of the AI ​​system among medical professionals.

[0025] S2.2: Extract acoustic features from the speech data, and based on the acoustic features, quantify the patient's anxiety and fatigue emotional state through speech emotion recognition technology. The acoustic features include pitch, speech rate, and energy. S2.3: Use OCR technology to extract numerical information from the image data. By automatically recognizing numerical values ​​in the medical device screen image through OCR technology, memory errors and transcription errors that may be caused by the patient's subjective description or manual input are completely avoided, ensuring the original accuracy and traceability of core physiological indicators such as blood pressure and blood sugar.

[0026] This embodiment performs sentiment analysis by integrating text semantic features with speech acoustic features (pitch, speech rate, and energy), which can accurately capture the patient's emotions. For example, it can judge anxiety by the trembling voice, thus achieving multidimensional and objective quantification of the patient's psychological state and overcoming the shortcomings of relying solely on text content for sentiment analysis, which is one-sided and inaccurate.

[0027] Furthermore, after feature extraction from the multimodal data, a multimodal data cross-validation step is included: comparing data of the same type from different modalities; if the difference exceeds a preset threshold, a follow-up inquiry procedure or manual verification process is initiated. In this embodiment, the data validation rules for hypertension patients are shown in Table 1: Table 1

[0028] The data validation rules for diabetic patients are shown in Table 2: Table 2

[0029] It should be noted that multimodal data cross-validation is triggered when patients upload the same data using different methods during follow-up. This cross-validation is not triggered if patients only upload text or images.

[0030] The core of chronic disease management lies in data-driven decision-making, and the accuracy of data directly determines the effectiveness and safety of interventions. This embodiment utilizes the redundancy between different modalities of data for mutual verification, effectively identifying and filtering abnormal data points caused by patient subjective misreporting, memory bias, measurement errors, or equipment malfunctions. For example, it successfully verifies the difference between the patient's stated blood pressure and the blood pressure in the device image, preventing high-risk misjudgments based on erroneous data, such as a patient misreporting 170 / 100 mmHg when the actual blood pressure is 120 / 80 mmHg, thus avoiding ineffective or even harmful interventions. Meanwhile, traditional data quality control heavily relies on manual judgment by medical staff, which is inefficient and difficult to scale. This embodiment, through preset configurable, clinically knowledge-based threshold rules, enables the system to automatically and in real-time detect data inconsistencies and intelligently trigger follow-up questioning processes. This constructs a closed-loop, self-correcting data acquisition system, reserving manual intervention only for the most complex abnormal situations, thereby significantly improving the system's operational efficiency.

[0031] In an optional embodiment, in step S3, the physiological risk score is calculated using the following formula:

[0032] in, For physiological risk scoring, Let i be the clinical weight of the i-th physiological indicator. Let i be the current measured value of the i-th indicator. Let be the clinical control target value for the i-th physiological indicator. The normalization function used to calculate the deviation between the current value and the target value is, for example, f(V, T) = |V - T| / T, and the result is normalized to the interval [0, 1]. In this embodiment, for hypertensive patients, the method for determining the clinical weights of physiological indicators and the source of the weights are shown in Table 3. Table 3

[0033] For diabetic patients, the methods for determining the clinical weights of physiological indicators and the sources of those weights are shown in Table 4: Table 4

[0034] Furthermore, the treatment adherence score is calculated using the following formula:

[0035] in, For treatment compliance scores, Medication adherence is calculated based on the ratio of actual to required doses. Dietary adherence is calculated based on the match between patient meal photos and recommended nutritional plans. Exercise adherence is calculated based on the match between patient exercise data and exercise prescription. , , The weighting coefficients are dynamically adjusted based on the patient's condition, and satisfy the following conditions: = 1. , , The value is not fixed, and for patients with hypertension... (medication) (diet), The baseline values ​​for (exercise) were 0.5, 0.3, and 0.2, respectively. When the patient's condition is severe or unstable, The weight of medication use increases when a patient's relevant metabolic indicators (such as blood lipids and blood sugar) are abnormal. The weight of diet increases when patients are at risk of related complications (such as obesity, cardiovascular problems). (Exercise) weight increases, but for diabetic patients, (medication) (diet), The baseline values ​​for (exercise) were 0.6, 0.25, and 0.15, respectively. The adjustment mechanism for each weight was also dynamically adjusted according to the patient's condition, and the adjustment mechanism was consistent with that for patients with hypertension. To avoid duplication, it will not be elaborated here. This dynamic weighting mechanism ensures that the assessment focus is consistent with the patient's most pressing health needs, achieving "one person, one plan" and dynamic optimization.

[0036] Furthermore, To calculate exercise adherence based on the match between patient exercise data and exercise prescription, three key exercise indicators are included: exercise duration, exercise intensity, and exercise frequency. First, the actual exercise duration is obtained. Actual exercise intensity Actual motion frequency Then extract the prescription exercise duration. Prescription exercise intensity Prescription exercise frequency Then, based on this data, the matching degree of each exercise metric is calculated, including the matching degree of exercise duration. The calculation method is as follows: =min( ) This formula ensures that the match rate does not exceed 1.0; that is, even if the actual value exceeds the prescribed value, it is still calculated as a perfect match. Exercise intensity match rate. The calculation method is as follows: =min( ) The formula for calculating motion frequency matching degree is as follows: =min( ) The average of the three matching degrees mentioned above is used to obtain the sports compliance score. The calculation method is as follows:

[0037] Based on sports compliance score get The calculation method is as follows: .

[0038] The final score is a value between 0 and 100. The higher the score, the better the match between the actual exercise and the prescription requirements.

[0039] Furthermore, Dietary adherence, calculated based on the match between patient meal images and recommended nutritional plans, also includes three key intake indicators: sodium, sugar, and calories. First, the actual sodium intake is obtained. Actual sugar intake and actual calorie intake Then extract the target sodium content from the recommended nutritional plan. Target sugar content and target calorie content Then, based on this data, the intake matching degree is calculated, including the matching degree of sodium content. The calculation formula is as follows:

[0040] Sugar content matching degree The calculation formula is as follows:

[0041] Calorie matching The calculation formula is as follows:

[0042] The closer the matching score is to 1, the closer the actual intake is to the target. The average of the three intake matching scores is then calculated to obtain the initial dietary score. The calculation method is as follows:

[0043] Finally, based on the initial diet score get The calculation method is as follows:

[0044] in, Ensure the score is between 0 and 1. It is a value from 1 to 100, with a higher score indicating a better match.

[0045] Lifestyle health score L score This score is used to assess a patient's overall lifestyle habits. It integrates data from multiple dimensions, including smoking volume, alcohol consumption, sleep quality (from wearable devices), and average daily steps. In this embodiment, the lifestyle health score L... score See Table 5:

[0046] To make the scoring more scientific, the system adopts an age-stratified standard. Table 6 shows an example of the scoring standard for the elderly group (≥60 years old): Table 6

[0047] Furthermore, the health status vector also includes an emotion score E obtained based on speech data recognition or text data analysis. sentiment In this embodiment, the emotion score is obtained through voice emotion recognition and text emotion analysis, with a value ranging from 0.0 to 1.0, and is used to quantify the patient's psychological state. sentiment The scoring system is shown in Table 7:

[0048] Finally, the above scores are combined with other key features to form a health state vector S. t : S t [ , L score E sentiment T since_last_contact ] Among them, T since_last_contact This represents the number of days since the last contact. In this embodiment, physiological risk score This transforms the abstract concept of "physiological risk" into a concrete, comparable numerical value. By introducing clinical weights *wi*, it reflects the different medical importance of various physiological indicators (such as blood pressure vs. BMI). The normalization function *f* eliminates the influence of the different dimensions of various indicators (such as blood pressure in mmHg and blood glucose in mmol / L), allowing them to be weighted and summed. This enables the algorithm to assess a patient's overall risk by comprehensively weighing multiple indicators, rather than viewing single data points in isolation, greatly improving the scientific rigor and accuracy of the status assessment. Treatment adherence scoring... This incorporates two major lifestyle factors, diet and exercise, into a unified quantitative framework. This is achieved through three independently calculable components (A...). med A diet A exer This allows for precise identification of specific aspects of poor patient adherence, such as: forgetting to take medication? Overeating? Lack of exercise? This provides the possibility for generating targeted interventions.

[0049] In an optional embodiment, in step S4, the reward is not based on the absolute value at a certain point in time, but rather on the positive change in the scores of each item in the state vector. The specific reward function is calculated using the following formula:

[0050] in, For the reward function, , , These represent the positive changes in the ratings at adjacent time points. This represents the improvement in the physiological risk score. If the risk decreases (Pscore decreases), this item is positive, and the model receives a positive reward. This represents the increase in treatment adherence score; if adherence improves, this item is positive, and the model receives a positive reward. The increase in lifestyle health score. = , = , = , Rate your lifestyle health. , , These are the weighting coefficients, and C is the interaction cost constant.

[0051] Furthermore, weighting coefficients , , The weighting is dynamically adjusted based on the patient's condition, which includes newly diagnosed patients, patients with unstable conditions, patients with poor compliance, or patients in a stable phase. For newly diagnosed patients, the weighting coefficient is adjusted accordingly. Greater than the weight coefficient and weighting coefficients For patients with unstable conditions, the weighting coefficient Greater than the weight coefficient and weighting coefficients For patients with poor compliance, the weighting coefficient Greater than the weight coefficient and weighting coefficients For patients in the stable phase, the weighting coefficient Greater than the weight coefficient and weighting coefficients In this embodiment, the rules for judging the patient's condition are shown in Table 8: Table 8

[0052] For example, for newly diagnosed patients, the focus should be on adherence, with the weight adjusted to... =0.3, =0.5, =0.2, for patients with unstable conditions, the focus should be on physiological risks, and the weight should be adjusted accordingly. =0.6, =0.2, =0.2, for patients with poor adherence, adherence needs to be emphasized, and the weight is adjusted to... =0.2, =0.6, =0.2, for patients in the stable phase, the weight is adjusted to =0.4, =0.3, =0.3.

[0053] Furthermore, C is an interaction cost constant, representing time cost or interaction cost. This incentivizes the model to achieve the same effect with fewer interventions, avoiding excessive disruption to the patient. In this embodiment, the rules for determining the value of C are shown in Table 9: Table 9

[0054] In this embodiment, by rewarding positive changes in various scores, the model is provided with dense and immediate reward signals. This accelerates the training convergence process of the reinforcement learning model and guides the model to focus on learning strategies that bring continuous improvement, rather than pursuing short-term gains.

[0055] In an optional embodiment, please refer to Figure 4 In step S5, the reinforcement learning action space defines all possible interventions that the model can output; this is a discrete set. Specifically, it includes: medication management: {sending medication reminders, inquiring about and recording reasons for missed doses, suggesting consultation with a doctor to adjust dosage}; nutritional advice: {pushing specific low-sodium recipes, sending food calorie lookup tools, suggesting reducing red meat intake and explaining the reasons}; exercise prescription: {suggesting 30 minutes of brisk walking, pushing videos of knee rehabilitation exercises, reminding to complete today's exercise goals}; health education: {pushing articles on hypertension complications, sending stress reduction and relaxation techniques}; emotional support: {sending encouraging messages, identifying negative emotions and suggesting communication with family or contacting a doctor}. Reinforcement learning can employ mature deep reinforcement learning algorithms, such as Deep Q-Networks (DQN) or its improved versions. The model completes initial training through millions of interactions with a "virtual patient" (a simulation environment built based on a large amount of historical EHR data), and then undergoes online fine-tuning and continuous learning in real-world applications. Finally, it outputs the optimal intervention action.

[0056] In an optional embodiment, please refer to Figure 5 In step S6, the optimal intervention action is converted into personalized natural language interactive content through a generative AI model.

[0057] For example, if the optimal intervention action output by reinforcement learning is as follows: Action = {action type 'nutrition advice ', detail: 'reduce sodium_intake_by_1g ',reason: 'P score_BP_high ', target user 'elderly male Using a large language model (LLM), combined with patient profiles (age, education level, occupation) and real-time emotional state from Esentiment scores, instructions are translated in a "humanized" way, specifically targeting elderly users (the target in the instructions). user 'elderly male( ) “Hello Mr. Wang, I’m your health assistant. The system noticed your blood pressure was a little high yesterday, don’t worry. When we cook today, try using a small spoonful less salt than usual, okay? This will help lower your blood pressure!” For younger, more educated users: “Hi [Username], the system detected your blood pressure is slightly higher than the target value. According to the latest health guidelines, it is recommended to reduce your daily sodium intake by about 1 gram. A simple change is to try using natural umami ingredients such as tomatoes and mushrooms to replace some of the salt when cooking. Click here to see some delicious low-sodium recipes.” In addition to text, LLM can also call other tools based on instructions to generate illustrated health summaries, simple charts showing blood pressure trends, or links to relevant health education videos, making communication more intuitive and engaging.

[0058] In this embodiment, the highly structured 'machine instructions' output by the reinforcement learning module are transformed into personalized natural language that is more human-like, easier to understand, and more acceptable. This can greatly improve patient acceptability and compliance.

[0059] Example 2 This embodiment, based on the method proposed in Embodiment 1, fully illustrates the entire process from data input to intervention plan generation using a patient with hypertension and a patient with diabetes. The basic information of the hypertension patient, Mr. Zhang, is shown in Table 10: Table 10

[0060] Multimodal data input (time t): Physiological data: Blood pressure 158 / 92 mmHg, heart rate 78 bpm. Behavioral data: Smart pillbox recorded 18 medication doses in 7 days (21 doses should have been taken); meal image analysis showed sodium content of approximately 2800 mg / day; wearable device recorded an average of 3200 steps per day. Interactive data: Slightly low voice tone; text included "worried" and "stressed".

[0061] Module Calculation Process: 1. Data Fusion and Verification: The difference between verbal blood pressure readings and device readings is within 5 mmHg (pass); the analysis results of dietary images do not match the patient's self-reported "bland diet" (marked as abnormal). 2. Physiological Risk Score (P score ): Combining the deviation and weight of each physiological indicator (systolic blood pressure weight 0.35, etc.), P is calculated. score = 0.313. 3. Treatment adherence score (A score Medication adherence: 18 / 21 = 85.7%; Dietary adherence: (based on sodium intake) ≈ 60%; Exercise adherence: (based on steps) ≈ 40%; Weighted calculation (medication weight 0.5, etc.): A score= 0.687.4. Lifestyle Score (L) score Based on the middle-aged group standard, L was calculated by weighting factors such as smoking (poor), drinking (good), and exercise (poor). score =47.5. 5. Emotional score (E) sentiment Based on a combination of voice and text analysis, the response was determined to be mildly negative. (E) sentiment = 0.30. 6. Final state vector (S) t ):[P score 0.313, A score 0.687, L score 47.5, E sentiment : 0.30].

[0062] Intervention decision and program generation, RL decision: Based on the current state vector, the model determines that the patient's main problems are poor lifestyle and dietary adherence. Therefore, "pushing low-sodium recipes" and "reminding the patient to complete today's exercise goals" are selected as priority actions.

[0063] Generative AI output: "Hello Mr. Zhang, I'm your health assistant. I noticed your blood pressure has been fluctuating recently, which may be related to eating a little too much salt. Here are some delicious and simple low-sodium recipes for you to try today! Also, the weather is nice this afternoon, remember to go for a walk and complete your goal of 6,000 steps, which is very helpful for lowering blood pressure." The basic information of diabetic patient Li is shown in Table 11: Table 11

[0064] Multimodal data input (time t): Physiological data: fasting blood glucose 8.8 mmol / L, 2-hour postprandial blood glucose 13.5 mmol / L. Behavioral data: insulin injection compliance 95%; dietary image analysis showed high carbohydrate intake. Interactive data: voice analysis indicated tension, with text containing phrases like "afraid" and "don't know what to do."

[0065] Module calculation process, data fusion and verification: Most data verifications passed. Physiological risk score (P score ): Combining indicators such as blood glucose and glycated glucose with their respective weights (glycated glucose weight 0.4, etc.), P is calculated. score = 0.381. Treatment adherence score (A) score ): Medication adherence is high (95%), but dietary adherence is average (65%). After weighted calculation (medication weight 0.6, etc.), A score = 0.826. Lifestyle Score (L) score The overall score is moderate. score= 55.0. Mood score (E) sentiment ): Determined as a newly diagnosed anxiety disorder, E sentiment = 0.25. Final state vector (S) t ):[P score 0.381,A score 0.826, L score 55.0, E sentiment : 0.25] Intervention decision and plan generation, RL decision: The system identified the patient as having a "new diagnosis" and exhibiting significant anxiety. Based on the reward weighting adjustment rules, prioritizing improving adherence (λ²=0.5) and providing emotional support, "pushing articles on basic diabetes knowledge" and "sending encouraging messages" were selected as priority actions.

[0066] Generative AI output: "Hello Ms. Li, it's normal to have some worries and questions when you first start learning about blood sugar management, so please don't feel pressured. We've prepared a simple and easy-to-understand article about diabetes to help you better understand it. You're doing a great job! Keep it up, and your blood sugar will definitely become more and more stable! Feel free to ask me any questions you may have." In this embodiment, the intervention strategies for Zhang and Li are drastically different. Zhang received specific behavioral instructions (low-sodium diet, exercise goals) due to lifestyle issues; while Li, newly diagnosed and anxious, received priority in knowledge dissemination and emotional support. This demonstrates that the method in Embodiment 1 can dynamically generate different intervention plans based on subtle differences in the patient's condition, which contrasts sharply with the static and uniform recommendation model in the prior art.

[0067] Example 3 Please see Figure 2 This invention also provides a chronic disease intervention program generation system based on reinforcement learning and multimodal data, comprising the following modules: Acquisition module: configured to acquire multimodal data of patients, including at least electronic medical record data, voice data, image data, text data, and physiological time-series data from wearable devices; Feature extraction module: configured to extract features from the multiple modal data to obtain multimodal feature vectors; The construction module is configured to construct a health status vector based on the multimodal feature vector, wherein the health status vector includes at least a physiological risk score, a treatment adherence score, and a lifestyle health score. Design module: configured to design a reward function based on the dynamic changes of the health state vector; Output module: configured to use a reinforcement learning algorithm to optimize a policy network in a predefined intervention action space based on the reward function, so as to output the optimal intervention action; The conversion module is configured to convert the optimal intervention into personalized natural language interactive content using a generative AI model.

[0068] The system proposed in this embodiment is based on the method proposed in Embodiment 1. Therefore, the options proposed in Embodiment 1 are also applicable to this embodiment. To avoid repetition, they will not be described again here.

[0069] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0070] The readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic storage (such as floppy disks, hard disks, magnetic tapes, magneto-optical disks (MO), etc.), optical storage (such as compact disks (CD), digital video discs (DVD), Blu-ray discs (BD), high-definition universal discs (HVD), etc.), and semiconductor storage (such as read-only memory (ROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), non-volatile memory (NAND FLASH), solid-state disks (SSD), etc.).

[0071] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for generating a chronic disease intervention plan based on reinforcement learning and multi-modal data, characterized in that, The method comprises the following steps: acquiring multi-modal data of a patient, the multi-modal data comprising at least electronic medical record data, voice data, image data, text data, and physiological time-series data from a wearable device; extracting features from the multi-modal data to obtain a multi-modal feature vector; constructing a health status vector based on the multi-modal feature vector, the health status vector comprising at least a physiological risk score, a treatment adherence score, and a lifestyle health degree score; designing a reward function according to dynamic changes in the health status vector; optimizing a policy network in a predefined intervention action space according to the reward function by using a reinforcement learning algorithm to output an optimal intervention action; converting the optimal intervention action into personalized natural language interactive content through a generative AI model.

2. The method of claim 1, wherein, The feature extraction from the multi-modal data specifically comprises: extracting medical entity, sentiment, and user intent features from the voice data and the text data by using a medical field pre-trained language model, the medical entity comprising one or more of symptoms, drugs, and physiological indicators; extracting acoustic features from the voice data, and quantifying the anxiety and fatigue emotional states of the patient based on the acoustic features through voice sentiment recognition technology, the acoustic features comprising pitch, speech rate, and energy; extracting numerical information from the image data using OCR technology.

3. The method of claim 1, wherein, The physiological risk score is calculated by the following formula: wherein, is a physiological risk score, is a clinical weight for the i-th physiological indicator, is a current measurement value for the i-th indicator, is a clinical control target value for the i-th physiological indicator, is a normalization function for calculating the degree of deviation of the current value from the target value.

4. The method according to any one of claims 1 to 3, characterized in that, The treatment adherence score is calculated by the following formula: Wherein, is a treatment adherence score, is a medication adherence calculated based on the ratio of actual and prescribed medication taking times, is a diet adherence calculated based on the matching degree of patient meal pictures and recommended nutritional regimen, is an exercise adherence calculated based on the matching degree of patient exercise data and exercise prescription; , , is a weight coefficient dynamically adjusted according to the patient's condition, and satisfies = 1.

5. The method of claim 4, wherein, The reward function is calculated by the following formula: wherein, is a reward function, , , are positive change amounts of the score at adjacent time points, respectively, and = , = , = , is a lifestyle health score, , , are weight coefficients, and , C is an interaction cost constant.

6. The method of claim 5, wherein, weighting factor , , According to the dynamic adjustment of the patient state, the patient state includes newly diagnosed patients, unstable patients, poor compliance patients or stable patients, and for newly diagnosed patients, the weighting factor is greater than the weighting factor and the weighting factor , for unstable patients, the weighting factor is greater than the weighting factor and the weighting factor , for poor compliance patients, the weighting factor is greater than the weighting factor and the weighting factor , for stable patients, the weighting factor is greater than the weighting factor and the weighting factor .

7. The method of claim 1, wherein, After the feature extraction from the multi-modal data, the method further comprises a multi-modal data cross-checking step: comparing the same type of data from different modalities, and if the difference exceeds a preset threshold, starting a follow-up question program or an artificial verification process.

8. The method of claim 1, wherein, The health status vector further comprises an emotional score obtained based on voice data recognition or text data analysis. 9.A system for generating a chronic disease intervention program based on reinforcement learning and multi-modal data, characterized in that, The system comprises the method according to any one of claims 1-8, comprising: a collection module configured to acquire multi-modal data of a patient, the multi-modal data comprising at least electronic medical record data, voice data, image data, text data, and physiological time-series data from a wearable device; a feature extraction module configured to extract features from the multi-modal data to obtain a multi-modal feature vector; a construction module configured to construct a health status vector based on the multi-modal feature vector, the health status vector comprising at least a physiological risk score, a treatment adherence score, and a lifestyle health degree score; a design module configured to design a reward function according to dynamic changes in the health status vector; an output module configured to optimize a policy network in a predefined intervention action space according to the reward function by using a reinforcement learning algorithm to output an optimal intervention action; a conversion module configured to convert the optimal intervention action into personalized natural language interactive content through a generative AI model.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the method according to any one of claims 1-8.

Images