Method for providing preventive medicine system and preventive medicine service

Abstract

To build an individual disease risk prediction model using a machine learning technique and prepare, update, and provide an individualized health intervention plan, to prevent a disease.SOLUTION: A preventive medicine system provide a preventive medicine service by using a disease risk prediction model 10 that has learned health diagnostic data and case data, a large-scale language model 20 that has learned a disease guideline and medical document and chart instruction recording, and a reinforcement learning model 30. The health diagnostic data of a user is input to the disease risk prediction model 10, and a predicted disease risk is presented to the user. When data of consultation or questions from the user is input to the large-scale language model 20, instructions and a plurality of behavioral change plans are presented to the user as a response from the large-scale language model 20. The data of the behavioral change plans and the like selected by the user is trained in the reinforcement learning model 30, and a result of training in the reinforcement learning model 30 is trained in the large-scale language model 20.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a preventive medicine system and a method for providing preventive medicine services. [Background technology]

[0002] The use of big data in the medical field in recent years has led to many research results, starting with descriptive epidemiological studies, such as the identification of new disease risk factors, identification of real-world drug effects and side effects, and the construction of disease risk prediction models. In particular, in the field of disease onset prediction, risk scores have been constructed using large-scale cohorts, primarily in Western countries (see, for example, Non-Patent Documents 1 and 2), and these have been used to stratify residents' disease onset risk, identify population segments that should receive medical intervention, and formulate disease guidelines and public health policies.

[0003] On the other hand, in Asia, there are no established, unique disease risk prediction models, and models developed in Western countries are currently being used. It has been pointed out that Asians have a different disease risk than Westerners due to differences in genetic predisposition and lifestyle, and there is a need to establish unique disease risk prediction models using machine learning techniques (see, for example, Non-Patent Documents 3 to 5). [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Piepoli MF, Hoes AW, Agewall S, Albus C, Brotons C, Catapano AL, et al., 2016 European Guidelines on cardiovascular disease prevention in clinical practice: The Sixth Joint Task Force of the European Society of Cardiology and Other Societies on Cardiovascular Disease Prevention in Clinical Practice (constituted by representatives of 10 societies and by invited experts): Developed with the special contribution of the European Association for Cardiovascular Prevention & Rehabilitation (EACPR), Eur J Prev Cardiol 2016 Jul;23(11):NP1-NP96. [Non-Patent Document 2] Goff DC Jr, Lloyd-Jones DM, Bennett G, et al; American College of Cardiology / American Heart Association Task Force on Practice Guidelines. 2013 ACC / AHA guideline on the assessment of cardiovascular risk: a report of the American College of Cardiology / American Heart Association Task Force on Practice Guidelines. J Am Coll Cardiol. 2014;63(25, pt B):2935-2959. [Non-Patent Document 3] Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W, et al., LightGBM: A Highly Efficient Gradient Boosting Decision Tree, Advances in Neural Information Processing Systems. 2017;30:3146-54. [Non-patent document 4] Akiba, T., Sano, S., Yanase, T., Ohta, T., and Koyama, M. (2019), Optuna: A next-generation hyperparameter optimization framework, KDD '19: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, July 2019, Pages 2623-2631. [Non-Patent Document 5] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, LoRA: Low-Rank Adaptation of Large Language Models, arXiv:2106.09685. Summary of the Invention [Problem to be solved by the invention]

[0005] In Japan, an accurate disease risk prediction model, which would be the first step in achieving cardiovascular disease prevention, has not been established. Therefore, even at the individual level, it is difficult to understand the specific risk of developing cardiovascular disease, and from a macro perspective, it is a major issue that the patient population that truly needs preventive intervention for cardiovascular disease has not been sufficiently identified.

[0006] Furthermore, even in cases where there is a high risk of developing cardiovascular disease, the risk factors and lifestyle habits that require intervention vary from person to person, but there is a lack of tools to provide tailored advice and suggestions based on clear scientific evidence.

[0007] The present invention has been made in consideration of the above-mentioned problems, and aims to provide a preventive medicine system and a method for providing preventive medicine services that build an individual disease risk prediction model using machine learning technology, and plan, update, and provide personalized health intervention plans to prevent disease. [Means for solving the problem]

[0008] The preventive medicine system of the present invention includes a server and a user terminal connected to the server via a network, and is a preventive medicine system for providing preventive medicine services using a disease risk prediction model, a large-scale language model, and a reinforcement learning model, wherein the disease risk prediction model is trained using health checkup data as training data and case data as correct labels, and the large-scale language model is trained using disease guidelines, medical literature, and medical chart guidance records. When the server receives a user's health checkup data from the user terminal, it inputs the health checkup data into the disease risk prediction model, obtains the disease risk predicted by the disease risk prediction model and sends it to the user terminal. When the server receives consultation or question data from the user terminal, it inputs the received data into the large-scale language model, obtains disease guidance, lifestyle guidance, and multiple behavioral modification plans for the user as responses from the large-scale language model and sends them to the user terminal. When the server receives from the user terminal data on a behavioral modification plan selected by the user from the multiple behavioral modification plans or data on the progress of the selected behavioral modification plan, it trains the received data using the reinforcement learning model, and trains the large-scale language model on the results of learning from the reinforcement learning model.

[0009] The method for providing preventive medicine services of the present invention is a method for providing preventive medicine services using a disease risk prediction model, a large-scale language model, and a reinforcement learning model, wherein the disease risk prediction model is trained using health checkup data as training data and case data as correct labels, and the large-scale language model is trained using disease guidelines, medical literature, and medical chart guidance records, and includes the steps of: upon receiving a user's health checkup data from the user terminal by a server connected to the user terminal via a network, inputting the health checkup data into the disease risk prediction model, obtaining a disease risk predicted by the disease risk prediction model, and transmitting the obtained disease risk to the user terminal; upon receiving consultation or question data from the user terminal, inputting the received data into the large-scale language model, obtaining disease guidance, lifestyle guidance, and multiple behavioral modification suggestions for the user as responses from the large-scale language model, and transmitting the obtained disease guidance and lifestyle guidance and multiple behavioral modification suggestions to the user terminal; and upon receiving from the user terminal data on a behavioral modification suggestion selected by the user from the multiple behavioral modification suggestions or the progress of achieving the selected behavioral modification suggestion, training the received data using the reinforcement learning model, and training the large-scale language model with the results of training by the reinforcement learning model. [Effects of the Invention]

[0010] According to the present invention, by providing a preventive medicine service by cooperating a disease risk prediction model, a large-scale language model, and a reinforcement learning model, it is possible to plan, update, and provide an individualized health intervention plan for a user, thereby preventing the user from getting sick. [Brief explanation of the drawings]

[0011] [Figure 1] 1 is a conceptual diagram of a preventive medicine system according to an embodiment of the present invention. [Figure 2] 1 is a configuration diagram of a preventive medicine system according to an embodiment of the present invention. [Figure 3] FIG. 2 is a block diagram of a user terminal. [Figure 4] FIG. 1 is a schematic diagram showing the process of extracting cases from a database and constructing a disease risk prediction model. [Figure 5A]1 is a graph showing the discrimination ability of the disease risk prediction model of this embodiment and a conventional disease risk prediction model for myocardial infarction. [Figure 5B] 1 is a graph showing the calibration capabilities of the disease risk prediction model of this embodiment and a conventional disease risk prediction model for myocardial infarction. [Figure 6] FIG. 1 is a schematic diagram showing an algorithm for identifying contributing risk factors. [Figure 7] FIG. 10 is a schematic diagram showing the setting of improvement target values ​​for contributing risk factors and the optimization of intensity gradients. [Figure 8] FIG. 1 is a schematic diagram illustrating the construction of a large-scale language model specialized for preventive medicine. [Figure 9] FIG. 1 is a schematic diagram illustrating optimization of user tutoring using a reinforcement learning model. DETAILED DESCRIPTION OF THE INVENTION

[0012] Modern advances in medical information technology have made it possible to collect and analyze large amounts of health data, facilitating early disease detection and risk prediction. However, optimizing individual risk assessment and intervention using this data requires advanced medical knowledge as well as algorithms and model-building techniques, but practical models have yet to be developed. Furthermore, disease prevention requires not only disease onset prediction but also the identification of factors contributing to individual risk and the subsequent action to improve the individual's condition. In particular, primary prevention targets who have not yet developed serious diseases often prioritize their own preferences due to a low awareness of the risk of disease and are unable to take preventive action. To increase the likelihood of behavioral change in such populations, the disease prevention pipeline formed by disease risk prediction, identification of contributing risk factors, setting improvement targets, and action planning must be optimized to suit individual preferences.

[0013] In an embodiment of the present invention, multiple machine learning technologies work together to build an integrated platform that provides a personalized disease prevention pipeline. A conceptual diagram of a preventive medicine system that realizes such a platform is shown in Figure 1. The preventive medicine system mainly consists of the following three components (1) to (3). (1) Building Health Digital Twins as a disease risk prediction model10 and using it to predict disease risk; (2) Building a large-scale language model 20 specialized for preventive medicine and using it to provide disease guidance and lifestyle advice to users; (3) Optimizing individualized instruction for users using a reinforcement learning model 30.

[0014] 2, the preventive medicine system 1000 of this embodiment includes a server 100 that performs machine learning and the like, and a user terminal 200 connected to the server 100 via a network N. The number of user terminals 200 connected to the server 100 is not limited.

[0015] The server 100 generates trained models (disease risk prediction model 10, large-scale language model 20, reinforcement learning model 30) by training the machine learning models.

[0016] The user terminal 200 is a general-purpose computer such as a personal computer (PC) or a mobile terminal such as a smartphone. As shown in Fig. 3, the user terminal 200 includes a processor 202, a memory 204, a storage 206, an input unit 208, a display 210, and an I / F 212.

[0017] The storage 206 stores application programs for utilizing the preventive medicine services (disease risk prediction, disease guidance, lifestyle guidance, etc.) provided by the server 100, and the processor 202 executes processing related to the preventive medicine services in accordance with the application programs. A user can use the preventive medicine services by connecting the user terminal 200 to the server 100 via the network N.

[0018] <Disease risk prediction model> As shown in Figure 4, the server 100 extracts training case data for (N-n1-n2) people who meet predetermined conditions from the health checkup data of N people in a predetermined database (for example, a database provided by DeSC Healthcare Co., Ltd.), inputs this training case data into a machine learning model as a learning dataset, and trains it to construct a disease risk prediction model 10.

[0019] Examples of health checkup data when constructing a disease risk prediction model10 for predicting the onset of cardiovascular disease include 15 items that are thought to contribute to improving diagnostic ability (age, sex, BMI, systolic blood pressure, diastolic blood pressure, LDL cholesterol, HDL cholesterol, triglycerides, fasting blood glucose, smoking status, drinking frequency, exercise frequency, use of antihypertensive medication, use of diabetes medication, use of dyslipidemia medication) and medical history.

[0020] To select subjects for analysis from N people, first exclude those with a specific medical history or missing data (a total of n1 people). Predetermined medical history includes, for example, cardiovascular disease and renal replacement therapy. Missing data refers to missing data on, for example, smoking status, drinking frequency, and exercise frequency. This leaves (N-n1) people as subjects for analysis.

[0021] Furthermore, among the (N-n1) people, the analysis subjects are narrowed down to those who developed a specific disease within the past five years or whose health status has been tracked for more than five years. Examples of specific diseases include myocardial infarction, stroke, heart failure, and atrial fibrillation. This removes n2 people from (N-n1), leaving (N-n1-n2) people as the subjects for machine learning analysis. For example, if N = 1.3 million, n1 = 350,000, and n2 = 650,000, case data for 300,000 people will ultimately be obtained.

[0022] A disease risk prediction model 10 that predicts the onset of a specific disease (e.g., cardiovascular disease) within five years can be constructed by inputting a training dataset consisting of health checkup data as training data and case data (onset disease, health condition) as correct labels into a machine learning model and training it. Here, to avoid overfitting, 20% of the training dataset is used as performance test data, and the remaining 80% is used as training data. Then, 10% of the training data is used for monitoring, and 90% is used for training.

[0023] A decision tree gradient boosting algorithm (e.g., Light GBM disclosed in Non-Patent Document 3) was used as a prediction algorithm in machine learning, and hyperparameter tuning was performed using a search algorithm based on Bayesian optimization (e.g., Optuna disclosed in Non-Patent Document 4) to optimize the parameters of the machine learning model.

[0024] The server 100 calculates the relative risk of the predicted disease risk compared with that of people of the same age and gender. Specifically, the server 100 calculates the average risk according to age and gender from the learning dataset, and obtains the ratio to the predicted disease risk as the relative risk. The server 100 also obtains a histogram of the distribution of health checkup data for people of the same age and gender for each item (BMI, systolic blood pressure, diastolic blood pressure, LDL cholesterol, HDL cholesterol, triglycerides, and fasting blood glucose) from the learning dataset, and obtains the position of the input user's health checkup data on the histogram.

[0025] When the preventive medicine service application software is launched on the user terminal 200, and the user's health checkup data is sent from the user terminal 200 to the server 100 and input into the disease risk prediction model 10, the disease risk prediction model 10 predicts the user's disease risk and the relative risk compared to people of the same age and sex, as shown in Fig. 1. The predicted disease risk and relative risk, as well as the position on a histogram for people of the same age and sex in the input health checkup data, are displayed on the display 210. For example, if the user's LDL cholesterol is 150 mg / dL and the LDL cholesterol histogram for people of the same age and sex as the user is distributed in the range of 50 to 200 mg / dL, the position of 150 mg / dL on the histogram is clearly displayed.

[0026] Figures 5A and 5B show the results of comparing disease risk prediction model 10 using Light GBM with a conventional disease risk prediction model using logistic regression analysis for myocardial infarction. Figure 5A shows that the c-index, an index for evaluating prediction accuracy, is 0.762 for disease risk prediction model 10, while it is 0.672 for the conventional disease risk prediction model, indicating that disease risk prediction model 10 exhibits high discrimination ability. Figure 5B also shows that the observed probability of disease risk prediction model 10 is closer to the predicted probability than the conventional disease risk prediction model, indicating high calibration ability.

[0027] From the parameters used in disease risk prediction by the disease risk prediction model 10, it is possible to identify contributing risk factors (for example, the top few contributing risk factors), which are elements that contribute to an increase or decrease in the risk. FIG. 6 schematically shows an algorithm for identifying contributing risk factors. As shown in FIG. 6, the server 100 replaces each input individual parameter (health check data) one by one with the population average value for the same age and sex (however, for smoking, drinking habit, and exercise habit, they are replaced with no smoking, no drinking, and exercise habit, respectively), and inputs the replaced population average value for each parameter into the disease risk prediction model 10 to obtain a predicted disease risk B.

[0028] By comparing the disease risk B predicted when a hypothetical intervention is performed by averaging the individual parameters in a population with the disease risk A predicted by directly inputting the individual parameters into the disease risk prediction model 10, it is possible to predict to what extent each parameter contributes to an increase or decrease in the disease risk. Based on the results of this comparison, the server 100 identifies, as contributing risk factors, parameters that significantly reduce risk when replaced with the population average value.

[0029] For example, consider a case where a female user is predicted to have a high risk of developing myocardial infarction within the next five years. In the algorithm shown in Figure 6, each individual parameter (BMI, systolic blood pressure, LDL cholesterol, etc.) is replaced one by one with the average value for women of the same age, and the risk of myocardial infarction is repeatedly predicted using the disease risk prediction model 10 (virtual intervention). For example, if the risk of myocardial infarction is lowest when systolic blood pressure is taken as the average value, systolic blood pressure is identified as the top contributing risk factor, and the difference from the actual predicted value is output. Similarly, parameters with the greatest risk reduction are output as the top contributing risk factors.

[0030] After identifying the contributing risk factors, the server 100 optimizes the intensity gradient by setting the improvement target value of the contributing risk factor based on the predicted disease risk A, disease guidelines, and medical literature, as shown in FIG. 7. Here, intensity corresponds to the amount by which the risk is corrected. For example, if the user's current weight is 90 kg, setting the target value to 50 kg corresponds to high intensity, and setting the target value to 80 kg corresponds to low intensity. Note that the improvement target value of the contributing risk factor can be changed as desired by operating the input unit 208 of the user terminal 200.

[0031] <Large-scale language model> In this embodiment, a large-scale language model 20 is used as a platform for providing health management and disease guidance. By learning accurate medical information using the large-scale language model 20, users can receive lifestyle counseling in a chat format using a user terminal 200, and individualized and specific recommendations can be made with the goal of reducing each individual's disease risk.

[0032] As shown in Fig. 8, the server 100 constructs a large-scale language model 20 trained mainly on two training datasets (disease guidelines / medical literature and medical chart instruction records). Here, the model is integrated using low-rank adaptation (LoRA) (see, for example, Non-Patent Document 5). Details of the two training datasets are as follows (i) and (ii).

[0033] (i) Disease guidelines and medical literature Through expert review, for example, guidelines for cardiovascular disease and related diseases that are cardiovascular risk factors, including hypertension, diabetes, and dyslipidemia, as well as primary prevention guidelines from Japan, Europe, and the United States, will be collected and restructured into a learnable format. Furthermore, by extracting the latest peer-reviewed papers that have not yet been reflected in the guidelines through literature searches, knowledge related to preventive medicine will be accumulated. While the medical field is experiencing fragmentation of specialties and information inflation, making it practically difficult for even preventive medicine experts to acquire all the knowledge in related fields, large-scale language models20 can continuously update their literature learning, making it possible to provide disease information and guidance based on cutting-edge evidence.

[0034] (ii) Medical record guidance records The above-mentioned medical knowledge based on literature provides a framework for providing disease guidance to users. However, literature knowledge alone is insufficient to provide more specific lifestyle guidance based on individual preferences. Because individual lifestyles vary widely, uniform guidelines and literature knowledge alone cannot provide guidance that flexibly addresses these differences, making behavioral change difficult to achieve. To address this issue, this embodiment uses lifestyle guidance records from medical records as a training dataset. Lifestyle guidance records from electronic medical records contain individual disease risk assessments and goals, detailed dietary and exercise histories, and lifestyle histories, as well as corresponding guidance from nurses, registered dietitians, and physical therapists. This allows for the learning of specific guidance patterns for individual cases. In this embodiment, a large-scale language model 20 is constructed by extracting over 20,000 past medical record records, including dietary and exercise guidance, through a multi-institutional collaborative study. This model is trained to provide lifestyle guidance that takes into account Japan's unique lifestyle and culture and the diversity of its people.

[0035] By using a large-scale language model 20 that integrates these learning resources, it is possible to provide lifestyle advice and specific behavioral change suggestions that are based on cutting-edge evidence and take into account the preferences and background of each individual user. Note that the large-scale language model 20 also includes models that learn documents and information related to health guidance for health checkups and are specialized for supporting health guidance.

[0036] The user terminal 200 is equipped with an interactive interface, allowing the user to ask questions and seek advice about disease-related and lifestyle-related issues by interacting with the large-scale language model 20 using the interactive interface. The large-scale language model 20 can provide optimal responses tailored to the individual's background by inputting the above-mentioned contributing risk factors and the intensity gradient of their improvement target values. In addition to medical explanations about contributing risk factors and predicted diseases, the large-scale language model 20 can also provide the user with lifestyle advice, including diet and exercise advice, which is central to primary prevention.

[0037] <Reinforcement learning model> As shown in FIG. 9, a reinforcement learning model 30 can be used to optimize individualized instruction for a user. For example, in this embodiment, a reinforcement learning algorithm based on multi-armed bandit theory is applied to optimizing health interventions. The multi-armed bandit problem is a problem of selecting one option from multiple options and maximizing the reward associated with that selection. In this problem, while it is considered best to select the option with the highest current experienced expected reward (information utilization) in the short term, it is necessary to make a comprehensive selection (information exploration) taking into account the possibility that there may be options with unknown high expected rewards. In this embodiment, a reinforcement learning algorithm that determines the balance between utilization and exploration is utilized in two stages.

[0038] In the first stage, multiple behavioral change plans based on the guidance prescription are presented to the user, and the user is prompted to select a target behavioral change plan. Specifically, the user selects one of the multiple behavioral change plans (A, B, C) displayed on the display 210 of the user terminal 200. The server 100 trains the user-selected behavioral change plan using the reinforcement learning model 30, so that from the next time onwards, the server 100 can present behavioral change plans that better suit the user's preferences.

[0039] In the second stage, the user is asked to report the achievement status (achieved or not) of the selected behavioral change plan. Specifically, the user selects either "achieved" or "not achieved" displayed on the display 210 of the user terminal 200. The server 100 trains the achievement status reported by the user using the reinforcement learning model 30, so that from the next time onwards, the server 100 can present the user with options that are likely to be achieved. The learning results of the reinforcement learning model 30 are output to the large-scale language model 20 for learning, and are reflected in the strength setting of the improvement target values ​​of the contributing risk factors.

[0040] In this way, by repeating the two-stage learning cycle and linking with other components in the behavior change system, it is possible to present achievable behavioral goals that suit the user's lifestyle and preferences, maximizing the effectiveness of lifestyle guidance.

[0041] <How preventive medicine services are provided> 1 and 9, the method for providing preventive medicine services in this embodiment includes the following steps. First, the server 100 inputs the user's health checkup data received from the user terminal 200 into the disease risk prediction model 10, obtains the disease risk predicted by the disease risk prediction model 10, and transmits it to the user terminal 200, while identifying individualized contributing risk factors (Step 1). Next, the server 100 sets initial improvement target values ​​for the contributing risk factors in consideration of the predicted disease risk, disease guidelines, and medical literature (Step 2). Next, the server 100 inputs the set improvement target values ​​for the contributing risk factors into the large-scale language model 20 (Step 3). When the server 100 receives consultation or question data from the user terminal 200, it inputs the data into the large-scale language model 20 and obtains disease guidance, lifestyle guidance, and multiple behavioral modification suggestions for the user as responses from the large-scale language model 20, which it then transmits to the user terminal 200 (Step 4). When the server 100 receives the user's selection data (behavioral change proposals or subsequent achievement status) from the user terminal 200, it trains the reinforcement learning model 30 with the selected data (step 5). Based on the training results of the reinforcement learning model 30, the server 100 optimizes the intensity gradient of the improvement target values ​​of the contributing risk factors and trains the large-scale language model 20, and also trains the training results of the reinforcement learning model 30 in the large-scale language model 20, thereby optimizing multiple behavioral change proposals to be presented to the user (step 6). The training results of the reinforcement learning model 30 are reflected in the guidance provided to the user in real time. As a result of this health intervention, the user's health checkup data is re-entered into the disease risk prediction model 10 at every health checkup, which is held every six months to one year, to continue disease prevention.

[0042] As shown in Figure 1, by using the preventive medicine service, a user experiences feedback in three cycles with different time intervals (short-term cycle, medium-term cycle, and long-term cycle). In the short-term cycle, the user interacts with the large-scale language model 20 to receive feedback on disease-related questions and lifestyle advice, including diet and exercise, and receive suggested behavioral change recommendations (white arrow). In the medium-term cycle, a behavioral change recommendation is selected, the intensity gradient of contributing risk factors is optimized according to the progress of the recommendation, and the recommendations are optimized by the large-scale language model 20 (dotted arrow). As a result, the user can receive optimal guidance and recommendations for behavioral change. In the long-term cycle, the user's health condition is reevaluated at each medical checkup, allowing the user to obtain more optimal disease risk prediction results (solid arrow).

[0043] As described above, the preventive medicine system 1000 of this embodiment includes multiple components: construction of a disease risk prediction model 10 (Health Digital Twins) based on health checkup data; identification of contributing risk factors; introduction of the concept of a strength gradient of improvement target values; construction of a large-scale language model 20 specialized for preventive medicine; and construction of a reinforcement learning model 30 for disease guidance. In the preventive medicine system 1000, these components are not independent but have mutual inputs and outputs, allowing them to operate as an organically linked system. This is a core concept for providing optimal interventions that always reflect the preferences of users who generally have difficulty taking preventive measures themselves. In this way, the preventive medicine system 1000 can utilize individual health data to provide innovative means for personalized disease prevention and health promotion.

[0044] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the spirit of the present invention. Other embodiments and modifications made by those skilled in the art are also included in the present invention. [Explanation of symbols]

[0045] 10 Disease Risk Prediction Models 20 Large-scale language models 30 Reinforcement Learning Models 100 servers 200 user terminals 1000 Preventive Medicine System

Claims

1. A preventive medicine system for providing preventive medicine services using a disease risk prediction model, a large-scale language model, and a reinforcement learning model, the system comprising: a server; and a user terminal connected to the server via a network, the preventive medicine system comprising: the disease risk prediction model is trained using health checkup data as training data and case data as correct labels, The large-scale language model is trained on disease guidelines, medical literature, and medical charts; The server When receiving the user's health checkup data from the user terminal, inputting the health checkup data into the disease risk prediction model, obtaining the disease risk predicted by the disease risk prediction model, and transmitting the disease risk predicted by the disease risk prediction model to the user terminal; When receiving consultation or question data from the user terminal, inputting the received data into the large-scale language model, obtaining disease guidance, lifestyle guidance, and multiple behavioral modification suggestions for the user as a response from the large-scale language model, and transmitting these to the user terminal; A preventive medicine system that, when it receives data from the user terminal regarding a behavioral change plan selected by the user from the plurality of behavioral change plans or the status of achievement of the selected behavioral change plan, trains the received data using the reinforcement learning model, and trains the learning results of the reinforcement learning model using the large-scale language model.

2. The server The health checkup data of the user is replaced with the average value of a group of people of the same age and sex as the user; By comparing the disease risk predicted by inputting the population average value into the disease risk prediction model with the disease risk predicted by directly inputting the health checkup data into the disease risk prediction model, contributing risk factors that contribute to an increase or decrease in the disease risk predicted by directly inputting the health checkup data are identified; setting a target value for improvement of the contributing risk factor based on the disease risk predicted by directly inputting the health checkup data, the disease guidelines, and medical literature; The preventive medicine system according to claim 1 , wherein the set improvement target value is learned by the large-scale language model.

3. 3. The preventive medicine system of claim 2, wherein the server optimizes an intensity gradient of the improvement target value corresponding to the correction amount of the contributing risk factor based on the learning results of the reinforcement learning model and trains the intensity gradient of the improvement target value corresponding to the correction amount of the contributing risk factor using the large-scale language model, and also trains the learning results of the reinforcement learning model using the large-scale language model to optimize the plurality of behavioral modification proposals to be presented to the user.

4. A method for providing preventive medicine services using a disease risk prediction model, a large-scale language model, and a reinforcement learning model, comprising: the disease risk prediction model is trained using health checkup data as training data and case data as correct labels, The large-scale language model is trained on disease guidelines, medical literature, and medical charts; A server connected to the user terminal via a network receiving the user's health checkup data from the user terminal, inputting the health checkup data into the disease risk prediction model, obtaining the disease risk predicted by the disease risk prediction model, and transmitting the disease risk predicted by the disease risk prediction model to the user terminal; receiving consultation or question data from the user terminal, inputting the received data into the large-scale language model, and obtaining disease guidance, lifestyle guidance, and multiple behavioral modification suggestions for the user as a response from the large-scale language model, and transmitting the results to the user terminal; receiving, from the user terminal, data on a behavioral change plan selected by the user from the plurality of behavioral change plans or data on the progress of the selected behavioral change plan, learning the received data using the reinforcement learning model, and learning the learning results of the reinforcement learning model using the large-scale language model; and methods for providing preventive medicine services, including:

Images