Diabetes management system, method and program product based on autonomous evolution large model
By constructing a large-scale multimodal autonomous evolution model, the shortcomings of the diabetes management system in terms of individual dynamic changes and multimodal data fusion were solved, enabling the generation of personalized health management suggestions and continuous optimization of the system, thereby improving the system's credibility and practicality.
Patent Information
- Application Number
- CN202510831204.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-20
- Publication Date
- 2025-11-07
AI Technical Summary
Existing intelligent diabetes management systems are unable to adapt to individual dynamic changes, lack multimodal data fusion capabilities, and lack adaptive learning capabilities, resulting in unstable management.
We employ a self-evolving large-scale model with multimodal modeling and adaptive learning. Through multimodal data collection and preprocessing, multimodal embedding and feature fusion, health management suggestion generation, evolutionary learning module, and credible explanation and visualization module, we construct a unified multimodal embedding space to achieve continuous optimization and causal explanation.
It has improved the intelligence level of diabetes management, enabled the generation and continuous optimization of personalized health management suggestions, and enhanced the credibility and practicality of the system.
Smart Images
Figure CN120913739A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to computer large model technology, and in particular to a diabetes management system, method and program product based on an autonomous evolution large model. BACKGROUND
[0002] Diabetes is a metabolic disease characterized by chronic hyperglycemia, and the number of patients continues to grow. It is difficult to manage, and long-term reliance on manual judgment by medical personnel cannot meet the actual needs. In recent years, the application of artificial intelligence in the medical field has been continuously promoted, especially the development of deep learning technology, which makes it possible to identify and intervene based on models.
[0003] However, the existing intelligent diabetes management system still has many shortcomings: (1) mainly relying on static knowledge base, unable to adapt to individual dynamic changes; (2) most of the model structures use direct migration of general language models, lacking structure-level optimization for health management tasks; (3) lack of multi-modal data fusion capability, unable to comprehensively process heterogeneous data such as images, texts, voices and physiological signals; (4) lack of continuous learning and adaptive ability, unstable in actual application.
[0004] Therefore, it is urgent to develop a large language model system with multi-modal modeling, adaptive learning and strong explanation to improve the intelligent level of diabetes management. SUMMARY
[0005] In view of the problems existing in the prior art, the purpose of the present application is to provide a diabetes health management system based on an autonomous evolution large model with multi-modal modeling, adaptive learning and strong explanation.
[0006] In order to achieve the above-mentioned purpose of the application, the present application provides the following technical solutions:
[0007] A diabetes health management system based on an autonomous evolution large model, comprising:
[0008] A multi-modal data acquisition and preprocessing module for acquiring multi-modal data related to diabetes in patients and preprocessing the data;
[0009] A multi-modal embedding and feature fusion module for uniformly embedding and encoding the multi-modal data and aligning and fusing them into a fusion feature;
[0010] A health management suggestion generation module for generating personalized health management suggestions based on the fusion feature;
[0011] an evolutionary learning module configured to evolve the health management recommendation generation module based on a reinforcement learning model, the reinforcement learning model taking the health management recommendation generation module as an agent, taking the fused features as states, taking the health management recommendations generated after the input fused features as actions, taking a weighted sum of the health management recommendation score and the health management recommendation modification requirement degree as a reward function, and setting a target function as:
[0012] J(0) = E[w(r).R(s,a)]
[0013] w(r) = a • s(0r)
[0014] wherein J(0) represents the target function, E[] represents an expectation, w(r) represents a reward function confidence weight adjustment function, R(s,a) represents a reward of performing an action a in a state s, a e (0,1] represents a global scaling coefficient, b represents a sensitivity coefficient, s() represents a Sigmoid function, and r e [0,1] represents the health management recommendation score;
[0015] a credible explanation and visualization module configured to provide a basis traceability and trend display of the health management recommendations.
[0016] Further, the multi-modal data includes modal data of images, voices, texts, and physiological signals, and embodies patient diet content, basic demographic data, and basic physiological data.
[0017] Further, the multi-modal embedding and feature fusion module specifically includes:
[0018] a multi-modal embedding unit configured to uniformly encode each modal data into an embedding space of the same dimension by using a unified embedding strategy;
[0019] a multi-modal fusion unit configured to fuse the encoded multi-modal data by using an attention fusion mechanism to obtain fused features.
[0020] Further, the health management recommendation generation module is specifically a neural network model.
[0021] Further, the reward function is specifically:
[0022] R(s,a) = p1r + p2I f
[0023] wherein p1 and p2 represent weight coefficients, I f represents the health management recommendation modification requirement degree, wherein when the modification requirement degree exceeds a preset threshold, I f = 0, otherwise I f = 1.
[0024] Further, the trusted explanation and visualization module specifically includes:
[0025] a cause-effect path generation unit for calculating a visualization cause-effect reasoning chain according to the attention weight of each modality data:
[0026] a trend visualization unit for calculating a health trend score of the fusion feature sequence of the multi-modal data in the current time window, generating a visualization health trend score curve, and showing the evolution trajectory of the patient health over time;
[0027] an individual comparison visualization unit for calculating the average value of the fusion features of the multi-modal data of all patients in the same population and visualizing it.
[0028] Further, the cause-effect reasoning chain is specifically:
[0029] CausePath={(m i ,att i )|att i =Attention(m i )}
[0030] In the formula, CausePath represents the cause-effect reasoning chain, m i represents the modality i data in the multi-modal data, att i represents the attention weight of the modality i data, and Attention(m i ) represents the calculation of the attention weight of m i .
[0031] Further, the health trend score is calculated according to the following formula:
[0032] f t =Trend(Z 1:t )
[0033] In the formula, f t represents the health trend score at time t, Z 1:t ={Z1,…,Z t} represents the fusion feature sequence of the multi-modal data from time point 1 to time t, and Trend() represents the health trend score function.
[0034] A diabetes health management method based on an autonomous evolution large model, comprising:
[0035] Collecting multi-modal data related to diabetes of a patient and pre-processing;
[0036] Uniformly embedding and encoding the multi-modal data and aligning and fusing them into a fusion feature;
[0037] Generate personalized health management suggestions based on fusion features;
[0038] Evolve the health management suggestion generation module based on a reinforcement learning model, which takes the health management suggestion generation module as an agent, the fusion features as the state, the health management suggestions generated after inputting the fusion features as the action, the weighted sum of the health management suggestion score and the health management suggestion modification requirement as the reward function, and the target function as:
[0039] J(θ)=E[w(r).R(s,a)]
[0040] w(r)=α·σ(βr)
[0041] In the formula, J(θ) represents the target function, E[] represents the expectation, w(r) represents the reward function confidence weight adjustment function, R(s,a) represents the reward of performing action a in state s, α∈(0,1] represents the global scaling coefficient, β represents the sensitivity coefficient, σ() represents the Sigmoid function, and r∈[0,1] represents the health management suggestion score.
[0042] Provide basis traceability and trend display of health management suggestions.
[0043] A computer program product comprising computer programs / instructions that, when executed by a processor, implement the above method.
[0044] Compared with the prior art, the present application has the following advantages: the present application uses a self-evolving large language model, constructs a unified multi-modal embedding space, and fuses image, voice, text and physiological sensor information to solve the problem of integrating heterogeneous health data and realize context-aware modeling. Furthermore, by introducing a human-machine co-training mechanism, a feedback path and a causal explanation module, the model can be continuously optimized and has causal explanation capability, improving the practicality and reliability of the system and overcoming the major limitations of existing AI medical systems in terms of explanation and generality. The functions of continuous learning, adaptive feedback and autonomous evolution form a "learning AI medical assistant" architecture. BRIEF DESCRIPTION OF DRAWINGS
[0045] Figure 1 is the system block diagram of the diabetes health management system based on the self-evolving large model provided by the present application;
[0046] Figure 2 is the flowchart of the diabetes health management method based on the self-evolving large model provided by the present application. DETAILED DESCRIPTION
[0047] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.
[0048] Example 1
[0049] This invention provides a diabetes health management system based on a self-evolutionary large-scale model, such as... Figure 1 As shown, it includes:
[0050] The multimodal data acquisition and preprocessing module is used to acquire and preprocess multimodal data related to diabetes from patients.
[0051] The multimodal embedding and feature fusion module is used to uniformly embed and encode multimodal data and align and fuse them into a single fused feature;
[0052] The health management suggestion generation module is used to generate personalized health management suggestions based on fusion features;
[0053] An evolutionary learning module is used to evolve the health management suggestion generation module based on a reinforcement learning model. The reinforcement learning model treats the health management suggestion generation module as an agent, uses fused features as the state, uses the health management suggestions generated after inputting the fused features as actions, and uses the weighted sum of the health management suggestion score and the degree of modification required for the health management suggestions as the reward function.
[0054] The credible explanation and visualization module is used to provide evidence tracing and trend display for health management recommendations.
[0055] The following is a brief introduction to each module.
[0056] (I) Multimodal Data Acquisition and Preprocessing Module
[0057] The multimodal data acquired by the multimodal data acquisition and preprocessing module includes modal data of images, voice, text, and physiological signals, reflecting the patient's dietary content, basic demographic data, and basic physiological data.
[0058] Multimodal data acquisition can be performed in the following ways:
[0059] (1) Patient's diet description and image acquisition.
[0060] By analyzing patients' descriptions and photographs of their food, the image recognition and natural language processing capabilities of a large-scale model can be used to identify the dietary content of diabetic patients. This part will demonstrate the large-scale model's ability to process multimodal data (images and text). The specific steps are as follows:
[0061] A. The patient uses a mobile phone to take photos of each meal, which allows them to identify the type and quantity of food.
[0062] B. The patient's description of the food (such as name, portion size, cooking method, etc.), so that the results of image recognition can be further confirmed and supplemented;
[0063] C. Intelligent plate heat calculation: using a quantitative intelligent plate, the food weight, heat and nutrient intake of diabetic patients can be accurately calculated;
[0064] (2) Basic physiological data collection.
[0065] Obtained by collecting wearable devices such as sports bands, smart shoes, medical instruments or manual input. Wearable devices such as sports bands and smart shoes have multiple sensors (such as accelerometers, gyroscopes, heart rate sensors, etc.) built-in, which can continuously monitor the patient's exercise, steps, heart rate, blood pressure, metabolism and other physiological indicators. The large model processes these data to accurately identify the patient's energy consumption throughout the day. These data will be used to dynamically monitor the patient's physical condition.
[0066] (3) Basic demographic data collection.
[0067] Basic demographic data includes education, age, gender, environment and other characteristics. These data are obtained through patient records or questionnaires to help the model understand the patient's basic background information. For example, education, age, gender, environment, etc.
[0068] Preprocessing of multi-modal data includes data processing such as deduplication, normalization, and missing data filling.
[0069] Through the above data collection and multi-modal processing steps, the system can comprehensively and accurately understand the diet and exercise of diabetic patients, providing reliable data support for subsequent health management and personalized treatment. This not only improves the processing capacity and accuracy of the large model, but also provides a more comprehensive health management plan for patients.
[0070] (ii) Multi-modal embedding and feature fusion module
[0071] The multi-modal embedding and feature fusion module specifically includes:
[0072] The multi-modal embedding unit is used to adopt a unified embedding strategy to uniformly encode each modality data into an embedding space of the same dimension:
[0073]
[0074] where I ∈ R H×W×C represents image modality data, represents feature-extracted speech modality data, T = t 14 , t2, …, t n represents text modality data, represents physiological signal modality data, zI Y T L respectively represent embedding vectors of image, speech, text, and physiological signal modal data, represents a unified embedding strategy function of corresponding modal data;
[0075] a multi-modal fusion unit for adopting an attention fusion mechanism to fuse the encoded multi-modal data to obtain fusion features;
[0076] Z=Align(z I ,z Y ,z T ,z L )
[0077] Align() represents a multi-modal attention fusion mechanism based on cross-attention.
[0078] (III) Health management suggestion generation module
[0079] The health management suggestion generation module is specifically a large language model based on a neural network model, such as an LLM decoder, which can generate personalized health management suggestions, such as reducing sweets intake, based on the fusion features Z. It should be noted that the health management suggestion is only a suggestion for the patient's own body management and can only be used as a reference, and cannot be used as the final diagnosis result of the disease. The specific disease diagnosis should be based on the judgment of professional doctors. The health management suggestion can provide a reference for the doctor's diagnosis and treatment.
[0080] (IV) Evolution learning module
[0081] The evolution learning module specifically adopts a reinforcement learning model to model the system as a Markov process (MDP):
[0082]
[0083] wherein:
[0084] state is specifically the fusion features Z;
[0085] action is specifically the health management suggestion output by the health management suggestion generation module when the fusion features Z are input;
[0086] reward function R(s,a): R(s,a)=ρ1r+ρ2I f , ρ1 and ρ2 represent weight coefficients, r∈[0,1] represents a health management suggestion score, I f represents the degree of modification required for the health management suggestion, and when the degree of modification required exceeds a preset threshold, If = 0, otherwise I f = 1.
[0087] Policy π θ (a|s): the probability of health management recommendations generated by the health management recommendation generation module.
[0088] P: state transition probability, in this system, each round of recommendation generation is considered as a single step MDP process, so there is no explicit modeling of state transition P, P = 1.
[0089] Specifically, the health management recommendation score can be the score of an expert, such as a professional doctor, or a trained score model, or a specially set score software system, etc. I f Indicates the degree of modification of health management recommendations, which is a binary indicator function, indicating the degree of acceptance of the model generated recommendations by the doctor. If the doctor only modifies no more than 10% of the characters in the generated recommendations (such as synonym replacement, order adjustment), it is assigned a value of 1; if the modification amplitude exceeds 10%, or is completely rewritten, it is assigned a value of 0. The system can quantify the modification amplitude by comparing the edit distance (such as Levenshtein Distance) between the original recommendation and the modified recommendation.
[0090] When training the reinforcement learning model, the objective function is set as:
[0091] J(θ) = E[w(r).R(s,a)]
[0092] w(r) = a • σ(βr)
[0093] Update the model by approximating the policy gradient (such as PPO):
[0094]
[0095] In the formula, J(θ) represents the objective function, E[] represents the expectation, w(r) represents the reward function confidence weight adjustment function, R(s,a) represents the reward of performing action a in state s, a e (0,1] represents the global scaling coefficient, used to adjust the maximum value of the weight, β represents the sensitivity coefficient, determines the response strength of the score change to the weight, σ() represents the Sigmoid function, defined as for mapping the score to the interval (0,1) to form a smooth, nonlinear weight change, is the advantage function estimate. This mechanism makes high-confidence feedback have a greater impact on policy update, improving learning efficiency and quality.
[0096] (Five) Credible interpretation and visualization module
[0097] The credible interpretation and visualization module specifically includes:
[0098] a cause path generation unit for calculating a visualized cause and effect reasoning chain (blood glucose trend image -> attention boost -> generate dietary recommendations) according to the attention weights of the modal data:
[0099] CausePath = {(m i , att i )|att i = Attention(m i )}
[0100] In the formula, CausePath represents the cause and effect reasoning chain, m i represents the modal i data (such as blood glucose image, speech segment, etc.) in the multi-modal data, att i represents the attention weight of the modal i data, and Attention(m i ) represents the calculation of the attention weight of m i . The reasoning chain reveals the main basis for the model to generate recommendations, which helps the clinician to understand the cause and effect basis of the model judgment and improves the credibility of the system. The cause and effect reasoning chain is constructed based on the fusion representation and indirectly participates in the strategy optimization process;
[0101] a trend visualization unit for calculating a health trend score of the fusion feature sequence of the multi-modal data in the current time window, generating a visualized health trend score curve, and displaying the evolution trajectory of the patient's health over time; the health trend score is calculated according to the following formula:
[0102] f t = Trend(Z 1:t )
[0103] In the formula, f t represents the health trend score at time t, which can be understood as the quantitative result of the model on the overall health risk or change trend at the current time, Z 1:t = {Z1,..., Z t} represents the fusion feature sequence of the multi-modal data from time point 1 to time t, and Trend() represents the health trend score function, which is used to capture the health evolution process (such as blood glucose trend rising, lifestyle change, etc.) of the individual in the continuous period, which can be realized by sliding window average, RNN, time series clustering, etc. The health trend score curve is consistent with the modal embedding on which the health management recommendation is generated, which can assist the doctor to understand the source of the recommendation, and is used as a feedback to affect the reinforcement learning reward function design, thereby forming a closed-loop learning structure;
[0104] an individual comparison visualization unit for calculating the average value of the fusion features of the multi-modal data of all patients in the same population and visualizing the same:
[0105]
[0106] Z i z i represents the fusion feature of the i-th user at a certain time point, and N is the number of sample people. It reflects the overall health status semantic representation of the user at the time. avg z represents the average value of the multi-modal fusion representation of all users in the same population (such as the same gender, age range or disease type) at a certain time, which is a central trend estimate of the population state.
[0107] The system of the present application can be widely applied to the fields of chronic disease follow-up, digital health management, remote consultation platform, etc., and has the following application advantages: reducing the burden of doctors, improving the efficiency and consistency of consultation; enhancing patient compliance and improving individual blood glucose control quality; providing a general structure reference for the landing of medical AI models.
[0108] The present application solves the problem of modeling difficulty in data scarce scenarios, and maintains accuracy through small sample and meta-learning structure. For example, in an internal test containing 52 real patients, the present application can still reach 86% in the accuracy of blood glucose control recommendation under the condition of using only 20 labeled samples, which is significantly better than the traditional BERT fine-tuning method (72%). In addition, in the preliminary pilot of a hospital in Jiangsu and a chronic disease management platform, the system processes the continuous data upload and suggestion generation feedback from 10 users within two weeks, and the acceptance rate of the doctor's suggestions is more than 90%. The system realizes the closed loop of data collection-representation modeling-consultation generation-expert feedback-model updating; the model has the characteristics of strong explainability, traceable generation results and natural human-computer interaction; the system is suitable for personalized and long-term chronic disease follow-up management, and is superior to the existing static platform.
[0109] The device provided by the embodiment of the present application can be used to execute the method provided by the embodiment one of the present application, and has the corresponding functions and advantages of the execution method.
[0110] It is worth noting that in the above embodiments, each unit and module included is only divided according to the function logic, but is not limited to the above division, as long as the corresponding function can be realized; in addition, the specific name of each functional unit is only for convenient mutual distinction, and does not limit the protection scope of the present application.
[0111] The embodiments described above are only illustrative, wherein the modules illustrated as separate components can or can not be physically separated, and the components displayed as modules can or can not be physical modules, i.e., can be located in one place or distributed to multiple network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment scheme. Those skilled in the art can clearly understand that each embodiment can be realized by means of software and the necessary general hardware platform, and of course, it can also be realized only by hardware, as long as the function or action can be realized.
[0112] Embodiment two
[0113] The embodiment provides a diabetes health management method based on an autonomous evolution large model, as shown in the following formula (I) : Figure 2 The embodiment provides a diabetes health management method based on an autonomous evolution large model, as shown in the following formula (I) :
[0114] Collecting multi-modal data related to diabetes of a patient and pre-processing the multi-modal data;
[0115] Uniformly embedding and encoding the multi-modal data and aligning and fusing the multi-modal data into a fusion feature;
[0116] Generating personalized health management suggestions based on the fusion feature;
[0117] Evolution of the health management suggestion generation module based on a reinforcement learning model, wherein the reinforcement learning model takes the health management suggestion generation module as an agent, takes the fusion feature as a state, takes the health management suggestion generated after inputting the fusion feature as an action, takes a weighted sum of a health management suggestion score and a health management suggestion modification degree as a reward function, and sets a target function as the following formula (II) :
[0118] J (θ) = E [w (r).R (s, a) ]
[0119] w (r) = a.σ (βr)
[0120] In the formula, J (θ) represents the target function, E [] represents the expectation, w (r) represents the reward function confidence weight adjustment function, R (s, a) represents the reward of performing the action a under the state s, a ∈ (0, 1] represents the global scaling coefficient, β represents the sensitivity coefficient, σ () represents the Sigmoid function, and r ∈ [0, 1] represents the health management suggestion score;
[0121] Providing basis traceability and trend display of the health management suggestions.
[0122] The method of the embodiment is one-to-one corresponding to the system of the above-mentioned embodiment one, and the details are referred to the embodiment one, and will not be described herein.
[0123] Embodiment three
[0124] The embodiment of the present application also provides a computer product, for example, an app on a mobile phone, a tablet computer, an installation program on a computer, etc., which comprises computer programs / instructions, and the computer programs / instructions are executed by a processor to realize the method in the second embodiment. The code for executing the computer executable program for operating the present application can be written in one or more programming languages or a combination thereof, and the programming languages include object-oriented programming languages, such as Java, Smalltalk, C++, and also include conventional procedural programming languages, such as the "C" language or similar programming languages. The program code can be executed completely on a user computer, partially on a user computer, as an independent software package, partially on a user computer and partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, connected through the Internet by using an Internet service provider).
[0125] It should be understood that the above embodiments and the description in the specification are only the principles, main features and advantages of the present application, and various changes and improvements can be made to the present application without departing from the spirit and scope of the present application, and these changes and improvements all fall within the protection scope of the present application.
Claims
1. A diabetes health management system based on an autonomous-evolution large model, characterized by, Comprise: A multi-modal data acquisition and preprocessing module for acquiring and preprocessing multi-modal data related to diabetes of a patient; A multi-modal embedding and feature fusion module for uniformly embedding and aligning the multi-modal data into a fusion feature; A health management suggestion generation module for generating personalized health management suggestions based on the fusion feature; An evolutionary learning module for evolving the health management suggestion generation module based on a reinforcement learning model, wherein the reinforcement learning model takes the health management suggestion generation module as an agent, takes the fusion feature as a state, takes a health management suggestion generated after inputting the fusion feature as an action, takes a weighted sum of a health management suggestion score and a health management suggestion modification degree as a reward function, and sets a target function as: J(θ) = E[w(r).R(s,a)] w(r) = a·σ(βr) wherein J(θ) represents the target function, E[] represents an expectation, w(r) represents a reward function confidence weight adjustment function, R(s,a) represents a reward of performing an action a under a state s, a ∈ (0,1] represents a global scaling coefficient, β represents a sensitivity coefficient, σ() represents a Sigmoid function, and r ∈ [0,1] represents a health management suggestion score; A credible explanation and visualization module for providing basis traceability and trend display of the health management suggestions.
2. The autonomous large model-based diabetes health management system of claim 1, wherein, The multi-modal data comprises modal data of images, voices, texts, and physiological signals, and embodies patient diet content, basic demographic data, and basic physiological data. 3.The autonomous large model-based diabetes health management system of claim 1, wherein, The multi-modal embedding and feature fusion module specifically comprises: A multi-modal embedding unit for uniformly encoding each modal data into an embedding space of the same dimension by using a uniform embedding strategy; A multi-modal fusion unit for fusing the encoded multi-modal data by using an attention fusion mechanism to obtain a fusion feature. 4.The autonomous large model-based diabetes health management system of claim 1, wherein, The health management suggestion generation module is specifically a neural network model. 5.The autonomous large model-based diabetes health management system of claim 1, wherein, The reward function is specifically: R(s, a) = p1r + p2I f wherein ρ1 and ρ2 represent weight coefficients, respectively, and I f represents the degree of modification of the health management recommendation, wherein, when the degree of modification exceeds a preset threshold, I f = 0, and otherwise I f = 1. 6.The autonomous large model-based diabetes health management system of claim 1, wherein, The credible explanation and visualization module specifically comprises: A causal path generation unit for calculating a visualized causal reasoning chain according to attention weights of the multi-modal data: A trend visualization unit for calculating a health trend score of a fusion feature sequence of the multi-modal data under a current time window to generate a visualized health trend score curve and display a health evolution trajectory of the patient over time; An individual comparison visualization unit for calculating an average value of fusion features of multi-modal data of all patients in the same population and visualizing the same. 7.The autonomous large model-based diabetes health management system of claim 6, wherein, The causal reasoning chain is specifically: CausePath = {(m i , att i )| att i = Attention(m i )} In the formula, CausePath represents the causal reasoning chain, m i Att represents modality i in multimodal data. i Attention(m) represents the attention weights for modality i. i ) indicates that for m i Calculate attention weights. 8.The autonomous large model-based diabetes health management system of claim 6, wherein, The health trend score is calculated according to the following formula: f t = Trend(Z 1:t ) where f t represents the health trend score at time t, Z 1:t = {Z1,..., Z t} represents the fused feature sequence of the multi-modal data from time point 1 to time t, and Trend() represents the health trend score function. 9.A method for diabetes health management based on an autonomous evolutionary large model, characterized by, Comprise: Acquiring and preprocessing multi-modal data related to diabetes of a patient; Uniformly embedding and aligning the multi-modal data into a fusion feature; Generating personalized health management suggestions based on the fusion feature; The health management suggestion generation module is evolved based on a reinforcement learning model, the reinforcement learning model taking the health management suggestion generation module as an agent, taking the fusion features as states, taking the health management suggestions generated after the input fusion features as actions, taking the weighted sum of the health management suggestion score and the health management suggestion modification degree as a reward function, and setting a target function as: J(θ)=E[w(r).R(s,a)] w(r)=α·σ(βr) In the formula, J(θ) represents the target function, E[] represents the expectation, w(r) represents the reward function confidence weight adjustment function, R(s,a) represents the reward of performing the action a under the state s, a∈(0,1] represents a global scaling coefficient, β represents a sensitivity coefficient, σ() represents a Sigmoid function, and r∈[0,1] represents the health management suggestion score. The basis for providing the health management suggestion is traced and a trend is displayed.
10. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the method in claim 9.