Chronic disease risk assessment and intervention strategy generation system based on data analysis

Abstract

The application provides a chronic disease risk assessment and intervention strategy generation system based on data analysis. The system collects multi-source heterogeneous information including clinical examination, behavior record, environmental data, etc., extracts key features using data fusion technology, and strengthens the time and space characteristics of the data through spatiotemporal weighted tensor decomposition method. Combined with causal reasoning technology, the system can accurately assess the risk of chronic diseases of individuals, eliminate confounding factors, and provide more reliable risk prediction. In addition, the system dynamically generates personalized intervention strategies through reinforcement learning algorithm, adjusts intervention measures according to real-time health data, and ensures accurate chronic disease management. The application has efficient risk prediction ability and personalized intervention scheme, which helps to improve the accuracy and effect of chronic disease management.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare, specifically to a system for generating chronic disease risk assessment and intervention strategies based on data analysis. Background Technology

[0002] Currently widely used chronic disease risk assessment models (such as the Framingham Risk Score) mainly rely on patients' static physiological indicators (such as blood pressure, blood glucose, and blood lipids), neglecting the dynamic changes of time, space, and environmental factors. These traditional methods often employ linear or simple nonlinear models, making it difficult to capture the complex high-order interactions between multi-source heterogeneous data (including clinical data, environmental monitoring, behavioral records, and medical text data). Furthermore, traditional methods fail to adequately extract semantic information from medical texts, resulting in the inability to fully utilize all available information during risk prediction, potentially leading to biased risk estimation.

[0003] Existing intervention strategies are typically based on fixed plans developed after initial risk assessments, lacking dynamic feedback mechanisms and making it difficult to adjust interventions in a timely manner to adapt to real-time changes in patients' health status. This static intervention not only ignores individual patient differences but also fails to integrate the impact of real-time environmental, spatiotemporal changes, and behavioral data on health status. Therefore, there is an urgent need for a risk assessment method that integrates multimodal data, captures spatiotemporal dynamic features, and reduces confounding interference through causal reasoning. Furthermore, it should incorporate dynamic decision-making techniques such as reinforcement learning to generate personalized, real-time feedback-optimized intervention strategies to achieve precise chronic disease management. Summary of the Invention

[0004] The purpose of this invention is to solve the aforementioned problems. To this end, the technical solution adopted by this invention is as follows:

[0005] A data-driven chronic disease risk assessment and intervention strategy generation system, comprising:

[0006] The multimodal data acquisition module is responsible for collecting multimodal patient data, including clinical examination data, environmental monitoring data, behavioral record data, and medical text data; and preprocessing the collected multimodal patient data to generate patient multimodal variables.

[0007] The multimodal data fusion module is responsible for sequentially performing feature extraction and dimensionality reduction, modality importance scoring, modality weight coefficient normalization, and multimodal feature fusion on multimodal variables to form a unified low-dimensional fusion feature.

[0008] The spatiotemporal feature enhancement module is responsible for extracting and enhancing temporal and spatial information from the unified low-dimensional fusion features output by the multimodal data fusion module. The fusion features are sequentially rearranged in terms of time and space dimensions, a weight tensor is constructed, a spatiotemporal weighted tensor decomposition model is established, and finally, the time factor matrix, spatial factor matrix S, feature factor matrix F, and core tensor are output. ;

[0009] The causal-enhanced chronic disease risk assessment module is responsible for enhancing the risk assessment of chronic diseases through causal reasoning techniques. It uses causal graphs and causal reasoning models to identify and eliminate confounding factors. The causal reasoning model is constructed through counterfactual reasoning and causal effect estimation, and finally outputs a chronic disease risk score.

[0010] The dynamic intervention strategy generation module uses reinforcement learning technology to dynamically generate intervention strategies by combining chronic disease risk scores and multimodal features.

[0011] Furthermore, the feature extraction and dimensionality reduction of the multimodal data fusion module employs a first deep learning model to map multimodal variables to the same low-dimensional latent space. The mapped features are represented as follows:

[0012]

[0013] in, For the model parameters corresponding to each modal data, The variable representing mode m is... The dimension of the common potential space.

[0014] Furthermore, the modal importance scoring is achieved by introducing a scoring function. To distinguish the contribution of each modality in the spatiotemporal feature enhancement module, the scoring function is a second deep learning model;

[0015] The importance score is calculated using the following formula: for each sample i and modality m, the original score is calculated. :

[0016]

[0017] in This represents the feature representation of the i-th sample in mode m.

[0018] The modal weight coefficient normalization is based on the original scores of each modality, which are normalized using the Softmax function to obtain the normalized weight coefficients corresponding to the modality m data of the i-th sample. The formula is as follows:

[0019]

[0020] in, Take all modes of sample i in sequence, and the weights of all modes of each sample satisfy the following: ;

[0021] The multimodal feature fusion is the process of representing the various modal features of sample i. The unified low-dimensional fusion feature is obtained by weighted summation according to the normalized weights, as shown in the following formula:

[0022]

[0023] Where N represents the total number of samples.

[0024] Furthermore, the temporal and spatial dimension rearrangement involves combining all the fused features. Rearranged into a spacetime tensor X in the spatial and temporal dimensions;

[0025] The constructed weight tensor W is used to emphasize the importance of key spatiotemporal locations, and its construction method is as follows:

[0026] 1) Calculate time weights: For each time point in the time dimension An exponential decay function is used, with more recent data receiving higher weight. The formula is as follows.

[0027]

[0028] in, Indicates the current time point, Indicates the first At a certain point in time, Indicates the time decay coefficient. This represents the time weight at the k-th time point;

[0029] 2) Calculate spatial weights: For each location in the spatial dimension According to the reference position The distance weight is calculated using the following formula.

[0030]

[0031] in, This represents the j-th spatial location. Indicates spatial distance. Indicates spatial attenuation coefficient This represents the spatial weight of the j-th spatial location;

[0032] 3) Calculate modal feature weights: Dynamically adjust the weights based on the importance of different modalities, and calculate the weight of each feature using a neural network, as shown in the following formula:

[0033]

[0034] in, This represents all sample data corresponding to mode m. The feature weights represent mode m;

[0035] 4) Combined Weights: The weights of time, space, and modal features are combined to obtain the value of each element in the final weight tensor, as shown in the following formula.

[0036]

[0037] in, The ( )th weight tensor W represents the weight tensor. ) elements.

[0038] Furthermore, the spatiotemporal weighted tensor decomposition model is established based on the spatiotemporal tensor X and weight tensor W constructed above. The spatiotemporal tensor X is decomposed into the product of the core tensor and three factor matrices, with a noise term added, as shown in the following formula.

[0039]

[0040] in, The core tensor is T, S and F are the time factor matrix, spatial factor matrix and eigenfactor matrix, respectively, and E is the residual tensor;

[0041] The spatiotemporal weighted tensor decomposition model is designed as a weighted reconstruction error plus a nuclear norm regularization term, as follows;

[0042]

[0043] in, Represents element-wise product. Denotes the Frobenius norm. The nuclear norm is used to encourage low-rank factor matrices. It is a regularization parameter;

[0044] The spatiotemporal weighted tensor decomposition model is optimized using the alternating least squares method. After optimization, the factor matrices T, S, and F, as well as the core tensor, are output. .

[0045] Furthermore, the construction of the causal graph is used to represent the causal relationships between the multimodal variables. Nodes in the causal graph represent variables, and edges represent the causal relationships between variables.

[0046] The causal graph is constructed based on the factor matrices T, S, and F output by the spatiotemporal feature enhancement module;

[0047] The causal reasoning model eliminates the influence of confounding factors through counterfactual reasoning and causal effect estimation.

[0048] The counterfactual reasoning involves hypothesizing whether intervention I will affect the patient's health status Y; the formula for counterfactual reasoning can be expressed as:

[0049]

[0050] in, This indicates the health status under intervention I. It is a causal reasoning model.

[0051] Furthermore, the causal effect estimation identifies the key factors that truly influence chronic disease risk by calculating the causal effects of different variables on chronic disease risk. Calculated using the following formula:

[0052]

[0053] in, This indicates the health status at the time of intervention I. This indicates the state of health without intervention;

[0054] The output of the causally enhanced chronic disease risk assessment module is a chronic disease risk score R, which represents the patient's chronic disease risk level, and the formula is as follows:

[0055]

[0056] Where g is the risk assessment function, which combines causal effects and health status. Calculate chronic disease risk scores.

[0057] Furthermore, the reinforcement learning model defines a state space B, an action space A, and a reward function Re.

[0058] The state space B represents the patient's physical state, including the causal risk assessment result R output by the causal enhanced chronic disease risk assessment module and the patient's multimodal data.

[0059]

[0060] The action space A represents the intervention measures that can be taken;

[0061] The reward function Re represents the effect of the intervention. The reward function is designed based on the improvement in the patient's physical condition; if the physical condition improves, a positive reward is given; if the physical condition does not improve, a negative reward is given. The reward function Re is expressed as follows:

[0062] Re(b,a)=Health Improvement(b,a)

[0063] Among them, b This indicates the patient's current physical condition, a A represents the action chosen based on the current physical condition.

[0064] Furthermore, the reinforcement learning algorithm employs the Q-learning algorithm to learn the optimal intervention strategy by finding the optimal strategy π, which maximizes the cumulative reward of taking action a in state b.

[0065]

[0066] Where α is the learning rate, γ is the discount factor, b′ is the next state, and a′ is the next action. This represents the Q-learning algorithm;

[0067] The optimal strategy is represented as follows:

[0068]

[0069] The output of the modality generated by the dynamic intervention strategy is the optimal strategy. The generated personalized intervention strategy.

[0070] Compared with the prior art, the advantages of the present invention are as follows:

[0071] This invention integrates multi-source data and combines spatiotemporal feature enhancement and causal reasoning techniques to enable the system to more comprehensively and dynamically identify and assess chronic disease risks, thereby improving the accuracy of chronic disease risk assessment. Through comprehensive and accurate assessment, it can help medical personnel better understand the health status of patients and make more scientific treatment decisions.

[0072] Traditional methods for chronic disease risk assessment often fail to effectively eliminate confounding biases among different variables, affecting the reliability of predictions. However, by employing causal graphical models and counterfactual reasoning, the system can identify and eliminate these confounding factors, ensuring the credibility of the risk assessment results. This causal-enhanced risk assessment can more accurately reveal the true impact of each factor on patient health, avoiding misjudgments common in traditional methods.

[0073] This invention generates personalized intervention strategies through reinforcement learning, which can be dynamically adjusted based on the patient's real-time health status and multimodal data. Compared to traditional fixed intervention programs, this flexible approach can more accurately address changes in the patient's health, improve intervention effectiveness, reduce the long-term risk of chronic diseases, and improve the patient's quality of life. Attached Figure Description

[0074] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0075] Figure 1 This is a flowchart illustrating the framework of the present invention;

[0076] Figure 2 This is a flowchart illustrating the feature fusion process of the present invention;

[0077] Figure 3 This is a schematic diagram illustrating the generation of spatiotemporal features according to the present invention;

[0078] Figure 4 A diagram is generated for the intervention strategy of this invention. Detailed Implementation

[0079] To achieve the above objectives, the present invention provides a data analysis-based system for generating chronic disease risk assessment and intervention strategies, comprising:

[0080] Module 1: Multimodal Data Acquisition Module. This module is responsible for collecting various types of patient data from multiple channels, including clinical examination data (such as blood glucose, blood lipids, etc.), environmental monitoring data (such as PM2.5 concentration, geographical location information), behavioral record data (such as step count, lifestyle habits), and medical text data (such as electronic medical records, doctor's notes).

[0081] The collected multimodal data underwent preprocessing, including cleaning, normalization, format alignment, and missing value handling, to generate multimodal variables, providing high-quality input for subsequent data fusion and modeling. The preprocessed multimodal variables are as follows:

[0082]

[0083] Where N represents the number of samples, d m The feature dimensions for each modality (m∈{c,e,b,t} correspond to clinical examination data, environmental monitoring data, behavioral record data, and medical text data, respectively) are respectively.

[0084] The collected multimodal patient data forms the foundation of the entire system, and the subsequent multi-source data fusion module will directly utilize this data for integration and processing.

[0085] Module 2, Multimodal Data Fusion Module, integrates multimodal patient data from different data sources to form a unified multidimensional data representation. The specific steps are as follows:

[0086] First, feature extraction and dimensionality reduction are performed on multimodal variables to ensure that spatiotemporal and modal information between data can be preserved and transmitted, and to enable data from different modalities to have a unified semantic expression. To this end, a first deep learning model is used to project the preprocessed features of each modality into the same low-dimensional latent space. This first deep learning model can be a fully connected layer, a convolutional neural network, or an autoencoder, etc. The mapped feature representation is as follows:

[0087]

[0088] in, For the model parameters corresponding to each modal data, The dimension of the common potential space.

[0089] Secondly, to differentiate the contributions of each modality to the downstream task, a scoring function is introduced. The importance of each modality is scored using a second deep learning model. For each sample i and modality m, a raw score is calculated. :

[0090]

[0091] in Rd is the feature representation of the i-th sample in mode m, and the scalar score reflects the relevance of the mode to risk prediction and intervention decision-making in the current sample.

[0092] Then, the scores of each modality are normalized using the Softmax function to obtain the normalized weight coefficients:

[0093]

[0094] Thus, the weights of all modalities satisfy the following for each sample: Furthermore, a higher weight indicates that the modality is more important in that sample.

[0095] Finally, the features of each modality are weighted and summed according to normalized weights to obtain the fused feature representation:

[0096]

[0097] The fused data is not only passed to the spatiotemporal feature enhancement module for deep feature extraction, but can also be directly used in some basic risk assessments, providing support for information unification of the entire system.

[0098] Module 3, the Spatiotemporal Feature Enhancement Module, primarily extracts and enhances temporal and spatial information from the unified low-dimensional fusion features output by the multimodal data fusion module, capturing the dynamic evolution and regional differences inherent in the data. This module employs a Spatiotemporal Weighted Tensor Decomposition (ST-WTD) model. By performing multidimensional decomposition and weighted reconstruction of the original data tensors, it achieves accurate extraction of spatiotemporal features, providing richer and more robust high-order feature inputs for subsequent causal risk assessment and dynamic intervention strategy generation. The implementation steps of this module are as follows:

[0099] First, the fused features are rearranged into a tensor X in the time and space dimensions. ,in For time dimension (e.g., by day, week). For spatial dimensions (such as different city areas, hospital departments). The number of data modal dimensions (in this embodiment) =4);

[0100] Secondly, a weight tensor W is constructed to emphasize the importance of key spatiotemporal locations. The construction method is as follows:

[0101] 1) Calculate time weights: For each time point in the time dimension An exponential decay function is used, with more recent data receiving higher weight. The formula is as follows.

[0102]

[0103] in, Indicates the current time point, Indicates the first At a certain point in time, Indicates the time decay coefficient (e.g.) =7 indicates that data within a week has a higher weight.

[0104] 2) Calculate spatial weights: For each location in the spatial dimension According to the reference position The distance weight is calculated using the following formula.

[0105]

[0106] in, This represents the j-th spatial location. Indicates spatial distance. Indicates the spatial attenuation coefficient (e.g.) =1km indicates that data within 1km of the patient has higher weight.

[0107] 3) Calculate modal feature weights: Dynamically adjust the weights based on the importance of different modalities, and calculate the weight of each feature using a neural network, as shown in the following formula:

[0108]

[0109] in, This represents the weight of mode m.

[0110] 4) Combined Weights: The weights of time, space, and modal features are combined to obtain the value of each element in the final weight tensor, as shown in the following formula.

[0111]

[0112] in, The ( )th weight tensor W represents the weight tensor. ) elements.

[0113] Then, the ST-WTD model is established and solved. Using the spatiotemporal tensor X and weight tensor W constructed above, a spatiotemporal weighted tensor decomposition model is established. This model assumes that the fused data tensor can be decomposed into the product of the core tensor and three factor matrices, plus a noise term, as shown in the following formula.

[0114]

[0115] in, Let T be the core tensor, S be the time factor matrix, S be the space factor matrix, and F be the characteristic factor matrix, respectively, and E be the residual tensor.

[0116] The spatiotemporal weighted tensor decomposition model is designed as a weighted reconstruction error plus a nuclear norm regularization term, as shown below.

[0117]

[0118] in, Represents element-wise product. Denotes the Frobenius norm. The nuclear norm is used to encourage low-rank factor matrices. It is the regularization parameter.

[0119] The spatiotemporal weighted tensor decomposition model is optimized using the alternating least squares method. After optimization, the output factor matrices T, S, and F, as well as the core tensor, are obtained. Together, they can form a high-order spatiotemporal feature representation that preserves the temporal dynamics, spatial distribution, and feature interaction information of the data.

[0120] Module 4, the Causal Enhancement Chronic Disease Risk Assessment Module, focuses on enhancing chronic disease risk assessment through causal reasoning techniques. This module utilizes causal graphs and causal reasoning models to identify and eliminate confounding factors, thereby more accurately assessing chronic disease risk and providing a foundation for generating subsequent dynamic intervention strategies.

[0121] First, a cause-effect graph is constructed to represent the causal relationships between different variables. For multimodal variables... In a cause-effect graph, nodes represent variables, and edges represent causal relationships between variables.

[0122] The causal graph is constructed based on the factor matrices T (time factor matrix), S (spatial factor matrix), and F (feature factor matrix) output by the spatiotemporal feature enhancement module. Specifically, the factor matrices reveal the relationships between time, space, and features, which are used to infer causal relationships between variables.

[0123] Suppose that a certain time factor in T is related to blood glucose levels, and a certain spatial factor in S is related to environmental pollution, then we can infer through causal reasoning whether environmental pollution has a causal effect on blood glucose levels.

[0124] The core of the causal reasoning model is to eliminate the influence of confounding factors through counterfactual reasoning and causal effect estimation.

[0125] Counterfactual reasoning involves hypothesizing whether a certain intervention I (such as increased physical activity) will affect a patient's health status Y (such as blood glucose levels). The goal of counterfactual reasoning is to estimate how a patient's health status would change without intervention.

[0126] The formula for counterfactual reasoning can be expressed as:

[0127]

[0128] in, This indicates the health status under intervention I. It is a causal reasoning model.

[0129] The causal effect estimation involves calculating the causal effects of different variables on the risk of chronic disease to identify which variables are the key factors that truly influence the risk of chronic disease. The causal effect can be calculated using the following formula:

[0130]

[0131] in, This indicates the health status at the time of intervention I. This indicates the state of health without intervention.

[0132] The output of module 4 is a chronic disease risk score R, which represents the patient's chronic disease risk level, and the formula is as follows:

[0133]

[0134] Where g is the risk assessment function, which combines causal effects and health status. Calculate the risk score.

[0135] Module 5, the dynamic intervention strategy generation module, generates dynamic and personalized intervention strategies based on the causal risk assessment results from Module 4. This module utilizes reinforcement learning (RL) technology, combined with the patient's chronic disease risk score and multimodal data, to dynamically adjust the intervention strategy to adapt to real-time changes in the patient's health status.

[0136] The core of the reinforcement learning model is to learn the optimal intervention strategy through interaction with the environment. In this embodiment, a state space B, an action space A, and a reward function Re are defined.

[0137] The state space B represents the patient's physical state, including the causal risk assessment result R output by module 4 and the patient's multimodal data.

[0138]

[0139] The action space A represents the intervention measures that can be taken, such as adjusting the drug dosage, increasing exercise, and changing the diet.

[0140] Reward function Re: Represents the effect of the intervention. The reward function is designed based on the improvement of the patient's physical condition. For example, if the patient's blood sugar level decreases, a positive reward is given; if the blood sugar level increases, a negative reward is given.

[0141] Re(b,a)=Health Improvement(b,a)

[0142] Where b represents the patient’s current physical state, and a represents the action selected based on the current physical state.

[0143] The reinforcement learning algorithm employs Q-learning to learn the optimal intervention policy. The goal of Q-learning is to find a policy π that maximizes the long-term cumulative reward of taking action a in state b.

[0144]

[0145] Where α is the learning rate, γ is the discount factor, b′ is the next state, and a′ is the next action.

[0146] The optimal strategy is represented as follows:

[0147]

[0148] The output of this module is the optimal strategy. The generated personalized intervention strategy dynamically adjusts the intervention measures based on the patient's real-time health status and environmental data.

[0149] This embodiment integrates multi-source data (clinical indicators, environmental parameters, behavioral data, and medical texts) and combines spatiotemporal feature enhancement and causal inference techniques. This enables the system to more comprehensively and dynamically identify and assess chronic disease risks, improving the accuracy of risk assessments for chronic diseases such as diabetes and cardiovascular disease. This comprehensive and precise assessment helps medical personnel better understand patients' health status, thereby making more scientific treatment decisions.

[0150] This embodiment utilizes a causal graph model and counterfactual reasoning to identify and eliminate confounding factors, ensuring the reliability of risk assessment results. This causally enhanced risk assessment more accurately reveals the true impact of various factors on patient health, avoiding misjudgments common in traditional methods.

[0151] This embodiment generates personalized intervention strategies through reinforcement learning, which can be dynamically adjusted based on the patient's real-time health status and multimodal data. Compared to traditional fixed intervention programs, this flexible intervention approach can more accurately address changes in the patient's health, improve intervention effectiveness, reduce the long-term risk of chronic diseases, and improve the patient's quality of life.

[0152] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A system for chronic disease risk assessment and intervention strategy generation based on data analysis, characterized in that, The system comprises: A multi-modal data acquisition module responsible for acquiring multi-modal patient data, including clinical examination data, environmental monitoring data, behavior record data, and medical text data; and preprocessing the acquired multi-modal patient data to generate patient multi-modal variables; A multi-modal data fusion module responsible for sequentially performing feature extraction and dimensionality reduction, modality importance scoring, modality weight coefficient normalization, and multi-modal feature fusion on the multi-modal variables to form unified low-dimensional fusion features; The spatio-temporal feature enhancement module is responsible for extracting and strengthening the time and space information from the unified low-dimensional fusion feature output by the multi-modal data fusion module; the fusion feature is sequentially subjected to time and space dimension rearrangement, weight tensor construction, and establishment of a spatio-temporal weighted tensor decomposition model, and finally outputs a time factor matrix, a space factor matrix S and a feature factor matrix F and a core tensor ; A causally enhanced chronic disease risk assessment module responsible for enhancing the risk assessment of chronic diseases through causal reasoning technology, identifying and eliminating confounding factors using a causal graph and a causal reasoning model constructed through counterfactual reasoning and causal effect estimation, and finally outputting a chronic disease risk score; A dynamic intervention strategy generation module that dynamically generates intervention strategies using a reinforcement learning model in combination with the chronic disease risk score and multi-modal features; The reinforcement learning model defines a state space B, an action space A, and a reward function Re, The state space B represents the patient's physical state, including the causal risk assessment result R output by the causally enhanced chronic disease risk assessment module and the patient's multi-modal data; The action space A represents the intervention measures that can be taken; The reward function Re represents the effectiveness of the intervention measures, and the design of the reward function is based on the improvement of the patient's physical state. If the physical state improves, a positive reward is given, and if the physical state does not improve, a negative reward is given; the reward function Re is represented as follows: Re(b,a)=Health Improvement(b,a) wherein b represents the current physical state of the patient, a A represents an action selected based on the current physical state; The causal graph is used to represent the causal relationships between the multi-modal variables. The nodes in the causal graph represent variables, and the edges represent the causal relationships between variables; The construction of the causal graph is based on the factor matrices T, S, and F output by the spatio-temporal feature enhancement module; The causal reasoning model eliminates the influence of confounding factors through counterfactual reasoning and causal effect estimation; The counterfactual reasoning assumes whether the intervention measure I will have an impact on the patient's health state Y; the formula of the counterfactual reasoning can be expressed as: wherein, represents the health status under intervention I, is the causal inference model.

2. The chronic disease risk assessment and intervention strategy generation system based on data analysis according to claim 1, characterized in that: The feature extraction and dimensionality reduction of the multi-modal data fusion module uses a first deep learning model to map the multi-modal variables into the same low-dimensional latent space. The mapped features are represented as: wherein, are model parameters for each modality data, denotes the corresponding variable of modality m, is the dimension of the common latent space.

3. The chronic disease risk assessment and intervention strategy generation system based on data analysis according to claim 2, characterized in that: The modality importance score is obtained by introducing a scoring function to distinguish the contribution of each modality in the spatio-temporal feature enhancement module, and the scoring function is a second deep learning model. The importance score is formulated as follows, for each sample i and modality m, the raw score is computed : wherein is a feature representation of the ith sample under modality m; The modal weight coefficient normalization is based on the original score of each modal, normalized by a Softmax function, to obtain the normalized weight coefficient corresponding to the modal m data of the i th sample , as follows: wherein, The weights of all the modalities of each sample satisfy ; The multi-modal feature fusion is to represent each modal feature of the sample i as The normalized weight is weighted and summed to obtain the unified low-dimensional fusion feature, and the formula is as follows: Where N represents the total number of samples.

4. The chronic disease risk assessment and intervention strategy generation system based on data analysis according to claim 1, characterized in that: The time and space dimension rearrangement is to rearrange the all fusion features In the spatial, temporal dimension, rearrange into a space-time tensor X; The constructed weight tensor W is used to emphasize the importance of key spatio-temporal positions, and the construction method is as follows: 1) Calculate time weight: for each time point on the time dimension , the recent data weight is higher, the formula is as follows, wherein denotes the current time point, denotes the kth time point, denotes the kth time point, denotes a time decay coefficient, denotes a time weight for the kth time point; 2) Compute spatial weights: for each location in the spatial dimension , compute a weight according to the distance to the reference location , formula as follows, wherein, denotes the jth spatial position, denotes the spatial distance, denotes the spatial attenuation coefficient denotes the spatial weight of the jth spatial position; 3) Calculate the modality feature weight: dynamically adjust according to the importance of different modalities, calculate the weight of each feature through a neural network, and the formula is as follows: wherein, represents the corresponding all sample data of the modality m, represents the feature weight of the modality m; 4) Combination of weights: combine the weights of time, space and modality features to get the value of each element in the final weight tensor, formula as follows, in, The ( )th weight tensor W represents the weight tensor. ) elements.

5. The chronic disease risk assessment and intervention strategy generation system based on data analysis according to claim 4, characterized in that, The establishment of the spatio-temporal weighted tensor decomposition model is based on the above-mentioned constructed spatio-temporal tensor X and weight tensor W, the spatio-temporal tensor X is decomposed into the product of core tensor and three factor matrices, and a noise term is added, formula as follows, wherein, is a core tensor, T, S and F are a time factor matrix, a space factor matrix and a feature factor matrix, respectively, and E is a residual tensor; The spatio-temporal weighted tensor decomposition model is designed to add a weighted reconstruction error plus a core norm regularization term, represented as follows; where, denotes element-wise multiplication, denotes the Frobenius norm, denotes the nuclear norm, to encourage low-rank factor matrices, is a regularization parameter; The spatio-temporal weighted tensor decomposition model is optimized by using an alternating least squares method. After optimization, factor matrices T, S and F and a core tensor are output .

6. The chronic disease risk assessment and intervention strategy generation system based on data analysis according to claim 1, characterized in that, The causal effect estimates are used to identify key factors that truly affect the risk of chronic disease by calculating the causal effect of different variables on the risk of chronic disease, the causal effect are calculated by the following equation: wherein, represents the health status at the time of intervention I, represents the health status in the absence of intervention; The output of the causal enhanced chronic disease risk assessment module is the chronic disease risk score R, which represents the patient's chronic disease risk level, formula as follows: where g is a risk assessment function that combines causal effects and health status Calculate a chronic disease risk score.

7. The chronic disease risk assessment and intervention strategy generation system based on data analysis according to claim 1, characterized in that, The reinforcement learning algorithm adopts Q-learning algorithm to learn the optimal intervention strategy, by finding the optimal strategy π, so that the cumulative reward of taking action a under state b is maximized; where a is the learning rate, g is the discount factor, b' is the next state, a' is the next action, denotes the Q-learning algorithm; The optimal strategy is represented as follows: The output of the dynamic intervention strategy generation modality is an optimal strategy The generated personalized intervention strategy.

Images