Analysis and prediction system based on disease data

Abstract

The invention discloses an analysis and prediction system based on disease data, which belongs to the technical field of disease data analysis and comprises a cross-modal distillation data fusion module, a time domain causal chain analysis module, a multi-level disease risk prediction module, a disease critical state real-time monitoring module, a self-adaptive intervention strategy generation module and a clinical collaboration and feedback optimization module. A time decay factor and a smoothing coefficient are introduced through the time domain causal chain analysis module, the causal strength between variables is updated in real time, nonlinear changes of disease progression are adapted, and individual risk scores are generated through the multi-level disease risk prediction module in combination with the dynamic weight and the time decay factor of the causal atlas. A spatial smoothing function is introduced to suppress the influence of an abnormal value, group risks are quantified through a modal mean value and a standard deviation, individual weights are dynamically adjusted through a feedback mechanism, the long-term stability of prediction is improved, causal strength and feature importance are used for joint modeling, and the problem that a traditional statistical model ignores a potential causal path is avoided.

Description

Technical Field

[0001] The present invention belongs to the technical field of disease data analysis, and specifically refers to an analysis and prediction system based on disease data. Background Art

[0002] With the rapid development of information technology and the rapid accumulation of medical data, disease prediction systems based on data analysis play an increasingly important role in public health management and clinical decision-making assistance. Traditional disease prediction methods mainly rely on statistical models and expert experience, and these methods have certain advantages in dealing with structured and small-scale data;

[0003] However, there are still certain defects in the existing disease data analysis and prediction systems. The causal analysis of the existing disease data analysis and prediction systems depends on static relevance, and it is difficult to capture dynamic causal relationships. There is a lack of modeling of the coupling relationship between multi-modal data, resulting in a lag or misjudgment in the update of the causal map. A single risk scoring model is difficult to take into account individual and group differences and is sensitive to outliers. The prediction depends on static feature weights and cannot dynamically adapt to the non-linear characteristics of disease progression. Therefore, an analysis and prediction system based on disease data is proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide an analysis and prediction system based on disease data to solve the problems raised in the above background art.

[0005] To achieve the above purpose, the present invention provides the following technical solutions: An analysis and prediction system based on disease data, including a cross-modal distillation data fusion module, a time-domain causal chain analysis module, a multi-level disease risk prediction module, a real-time monitoring module for disease critical states, an adaptive intervention strategy generation module, and a clinical collaboration and feedback optimization module;

[0006] The cross-modal distillation data fusion module is used to integrate multi-source heterogeneous data and standardize it, and generate a lightweight fusion data set of multi-modal time series data through knowledge distillation;

[0007] The time-domain causal chain analysis module is used to obtain the dynamic causal relationship between variables by mining multi-modal time series data, and construct and update the causal map;

[0008] The multi-level disease risk prediction module is used to combine the multi-modal time series data of the cross-modal distillation data fusion module and the causal map of the time-domain causal chain analysis module to hierarchically predict the disease risk scores of individuals and groups;

[0009] The real-time monitoring module for disease critical states is used to obtain real-time vital sign data and the risk scores of the multi-level disease risk prediction module, and dynamically adjust the critical threshold to trigger an alarm;

[0010] The adaptive intervention strategy generation module is used to generate personalized intervention plans based on the risk scores of the multi-level disease risk prediction module and the critical state data of the disease critical state real-time monitoring module, and optimize the strategy through reinforcement learning;

[0011] The clinical collaboration and feedback optimization module is used to collect feedback data from doctors and patients on the adaptive intervention strategy generation module, and to update the cross-modal distillation data fusion module through federated learning to form a closed-loop optimization.

[0012] Preferably, the cross-modal distillation data fusion module is wirelessly connected to the multi-level disease risk prediction module and the time-domain causal chain analysis module, the time-domain causal chain analysis module is wirelessly connected to the multi-level disease risk prediction module, the time-domain causal chain analysis module and the multi-level disease risk prediction module are wirelessly connected to the disease critical state real-time monitoring module, the disease critical state real-time monitoring module and the multi-level disease risk prediction module are wirelessly connected to the adaptive intervention strategy generation module, the adaptive intervention strategy generation module is wirelessly connected to the clinical collaboration and feedback optimization module, and the clinical collaboration feedback optimization module is wirelessly connected to the cross-modal distillation data fusion module.

[0013] Preferably, the cross-modal distillation data fusion module is used to integrate and standardize multi-source heterogeneous data, generate lightweight fusion data set multi-modal time series data through knowledge distillation, collect raw data from different modalities, unify the format and annotate timestamps, remove noise, fill missing values, normalize numerical data, and encode categorical data;

[0014] Through knowledge distillation, it is compressed into a lightweight tensor form, timestamp alignment is completed, and the compressed features are combined with the original data labels to form a structured lightweight fusion data set with multimodal time series data;

[0015] It should be understood that knowledge distillation is divided into a teacher model and a student model. The teacher model performs end-to-end training on the original multi-source heterogeneous data to learn the potential correlations and temporal features in the data. The student model learns the output and intermediate features of the teacher model through distillation, and compresses the data into tensor form.

[0016] Preferably, the time domain causal chain analysis module, the step of mining the dynamic causal relationship between variables includes: obtaining multimodal time series data of the cross-modal distillation data fusion module;

[0017] It should be understood that, assuming that the time series data of each mode M is , the m modal index is expressed as , let the causal function , measuring variables right The causal strength of the variable i and j at time t is calculated.

[0018] Specifically, the formula for implementing dynamic causal weights is as follows:

[0019] ,

[0020] In the formula, represents the dynamic causal weight between variables i and j at time t, T represents the length of the time window, represents the time window index, represents the time decay factor, m represents the modality index, and M represents the total number of modalities, represents the coupling coefficient of modality m, , reflecting the contribution weight of the m-th modality to the causal relationship, represents the causal function, measuring the causal strength of variable to .

[0021] Preferably, for the time-domain causal chain analysis module, the steps for dynamically updating the causal strength include: The dynamic causal weight is the causal strength at the current time step. According to , calculate the dynamic causal strength from variable i to j , let the historical causal strength be , combine the current causal weight and the historical causal strength , through the smoothing coefficient balance the influence of the current weight and the historical causal strength. The dynamic update of the causal strength is achieved as follows:

[0022] ,

[0023] In the formula, represents the dynamic causal strength, the dynamic causal strength from variable i to j, represents the smoothing coefficient, represents the non-linear activation function, expressed as , compress the causal strength to the range (0, 1), simulating the non-linear response of the biological system to the causal relationship, The final causal strength is based on the weighted average of the current causal weight and the historical causal strength, and recursive correction is used to avoid misjudgment of the causal relationship caused by short-term noise, represents the historical causal strength.

[0024] Preferably, for the time-domain causal chain analysis module, let the causal graph be , the steps for updating the topology of the causal graph include: According to the dynamic causal strength , set the dynamic threshold and the node importance , through the dynamic threshold And node importance The dynamic adjustment of the existence of edges and the topological update of the causal graph are implemented as follows:

[0025] ,

[0026] In the formula, represents the causal graph at time t, represents the variable 's causal edge, represents the dynamic causal strength, represents the dynamic threshold, represents the node importance, represents the maximum importance value of all nodes at the current moment. High-importance nodes are close to , and the threshold of the causal edge decreases. The dynamic threshold is dynamically adjusted according to the causal strength distribution.

[0027] The time-domain causal chain analysis module provides, through the multimodal coupling coefficient and the node importance adjustment mechanism, for the risk prediction module:

[0028] Individual level: The personalized risk factor weights of patients; Population level: The dynamic propagation path of regional epidemics. The causal chain analysis module corrects the errors caused by outdated data or noise interference in the multi-level disease risk prediction module by dynamically updating the causal graph.

[0029] Preferably, for the multi-level disease risk prediction module, the individual disease risk scoring steps include: According to the multimodal time-series data of the cross-modal distillation data fusion module and the causal graph of the time-series causal chain analysis module the causal strength in the feature importance of, combined with determining the dynamic weight of individual i to variable j, the dynamic risk score of individual i is implemented as:

[0030] ,

[0031] In the formula, represents the individual dynamic risk score, the dynamic risk score of individual i at time t, d represents the total number of variables, represents the latest update time of variable j, represents the time decay factor, represents the dynamic weight of individual i to variable j, is represented as , , represents the hyperparameter, balancing the contributions of causal strength and feature importance, represents through the multimodal time-series data Calculate the feature importance representing the current risk contribution of the quantization variable j

[0032] Preferably, for the multi-level disease risk prediction module, the population risk scoring step includes: setting the population characteristics as , according to the individual dynamic risk score Combined with the population characteristics as , calculate the risk score of population p, implemented as:

[0033] ,

[0034] In the formula, represents the population risk score, the risk score of population p at time t represents the number of individuals in population p , represents the mean and standard deviation of population p on modality m represents the spatial smoothing function is represented as , suppressing the influence of outliers on the population score represents the population characteristic weighting factor

[0035] Preferably, for the multi-level disease risk prediction module, the risk score feedback optimization step: according to the population risk score , adjust the weight of the individual dynamic risk score through the feedback mechanism , the risk score feedback optimization is implemented as:

[0036] ,

[0037] In the formula, represents the updated weight of individual i for variable j at time t + 1 represents the learning rate represents the dynamic weight of individual i for variable j represents the adjustment increment of the weight of individual i for variable j at time t is represented as , represents the feedback intensity coefficient represents the gradient of the population risk score with respect to the individual weight

[0038] Specifically, the multi-level disease risk prediction module utilizes the time decay factor and smoothing coefficient in the causal graph to improve the stability and accuracy of the prediction;

[0039] The individual dynamic risks obtained by the multi-level disease risk prediction module are integrated into the population risk data for a more perfect and stable population risk score. The individual data is updated and corrected according to the population risk score to improve the accuracy.

[0040] It should be understood that the multi-level disease risk prediction module is adjusted according to the dynamic causal strength in the causal map. When the causal strength of a certain disease in the causal map increases, the weight of this variable in the individual model is significantly higher than other variables. When the dynamic causal strength in the causal map is updated due to new data, the risk prediction module will automatically recalculate the individual and population scores. Over time, the strength of the causal relationship may gradually increase or decrease, and the risk prediction module maintains the timeliness of the prediction through dynamic weight adjustment.

[0041] Preferably, the disease critical state real-time monitoring module collects vital sign data in real time through a monitoring device, and inputs the real-time collected vital sign data into the multi-level disease risk prediction module for real-time risk scoring;

[0042] It should be understood that according to the patient's historical vital sign data and risk score, a dynamic threshold is calculated, and the threshold is dynamically adjusted according to the real-time risk score distribution. According to the real-time individual dynamic risk score The degree of deviation from the threshold is used to divide the warning level. When the real-time individual dynamic risk score exceeds the dynamic threshold, it is determined that the patient has entered the critical state.

[0043] Preferably, the adaptive intervention strategy generation module obtains the risk score of the multi-level disease risk prediction module and the critical state data of the disease critical state real-time monitoring module, and integrates the risk score, critical state data, and patient characteristics into a unified strategy input vector;

[0044] It should be understood that according to clinical guidelines and expert experience, basic intervention rules are defined, rule adaptation is carried out in combination with the patient's individual characteristics, the rule library is matched according to the current state, candidate intervention plans are generated, the intervention strategy generation problem is modeled as a Markov decision process, the policy network is trained through a deep reinforcement learning algorithm, and the reward function weight is dynamically updated. The current state is input into the trained policy network to output the optimal intervention action.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] 1. The present invention introduces a time decay factor and a smoothing coefficient through the time-domain causal chain analysis module to update the causal strength between variables in real time, adapts to the non-linear changes in disease progression, quantifies the contribution of different modalities to the causal relationship through the modal coupling coefficient, constructs a more accurate causal map, and has a dynamic threshold and node importance adjustment mechanism to avoid short-term noise interference and ensure the stability of causal chain analysis;

[0047] 2. The present invention generates an individual risk score by combining the dynamic weights and time decay factors of a causal graph in a multi-level disease risk prediction module, introduces a spatial smoothing function to suppress the influence of outliers, quantifies the population risk through the modal mean and standard deviation, dynamically adjusts the individual weights through a feedback mechanism to enhance the long-term stability of the prediction, jointly models using causal strength and feature importance to avoid the problem of traditional statistical models ignoring potential causal paths, integrates the individual dynamic risk into the population risk data for a more perfect and stable population risk score, updates and corrects the individual data based on the population risk score to improve accuracy, and maintains the timeliness of the prediction through dynamic weight adjustment;

[0048] 3. The present invention dynamically modifies the threshold based on historical data and real-time risk score distribution through a disease critical state real-time monitoring module to enhance the warning sensitivity, combines the risk score with vital sign data to avoid misjudgment by a single indicator, and divides the warning level according to the degree of deviation from the threshold to reduce waste of medical resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a schematic structural diagram of the analysis and prediction system based on disease data of the present invention;

[0050] Figure 2 is the operation flow of the analysis and prediction system based on disease data of the present invention Figure 1 ;

[0051] Figure 3 is the operation flow of the analysis and prediction system based on disease data of the present invention Figure 2 ;

[0052] Figure 4 is the operation flow of the analysis and prediction system based on disease data of the present invention Figure 3 . DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0054] Embodiment

[0055] Please refer to Figures 1-4 as shown. The present invention provides a technical solution: including a cross-modal distilled data fusion module, a time-domain causal chain analysis module, a multi-level disease risk prediction module, a disease critical state real-time monitoring module, an adaptive intervention strategy generation module, and a clinical collaboration and feedback optimization module;

[0056] The cross-modal distillation data fusion module is used to integrate multi-source heterogeneous data and standardize it, and generate a lightweight fusion dataset of multi-modal time-series data through knowledge distillation;

[0057] The time-domain causal chain parsing module is used to obtain the dynamic causal relationships between variables by mining multi-modal time-series data, and construct and update the causal graph;

[0058] The multi-level disease risk prediction module is used to combine the multi-modal time-series data of the cross-modal distillation data fusion module and the causal graph of the time-domain causal chain parsing module to hierarchically predict the disease risk scores of individuals and groups;

[0059] The disease critical state real-time monitoring module is used to obtain real-time vital sign data and the risk scores of the multi-level disease risk prediction module, and dynamically adjust the critical threshold to trigger an alarm;

[0060] The adaptive intervention strategy generation module is used to generate a personalized intervention plan according to the risk scores of the multi-level disease risk prediction module and the critical state data of the disease critical state real-time monitoring module, and optimize the strategy through reinforcement learning;

[0061] The clinical collaboration and feedback optimization module is used to collect feedback data from doctors and patients on the adaptive intervention strategy generation module, and update the cross-modal distillation data fusion module through federated learning to form a closed-loop optimization.

[0062] In this embodiment, optionally, the cross-modal distillation data fusion module is wirelessly connected to the multi-level disease risk prediction module and the time-domain causal chain parsing module, the time-domain causal chain parsing module is wirelessly connected to the multi-level disease risk prediction module, the time-domain causal chain parsing module and the multi-level disease risk prediction module are wirelessly connected to the disease critical state real-time monitoring module, the disease critical state real-time monitoring module and the multi-level disease risk prediction module are wirelessly connected to the adaptive intervention strategy generation module, the adaptive intervention strategy generation module is wirelessly connected to the clinical collaboration and feedback optimization module, and the clinical collaboration feedback optimization module is wirelessly connected to the cross-modal distillation data fusion module.

[0063] Preferably, the cross-modal distillation data fusion module is used to integrate multi-source heterogeneous data and standardize it, generate a lightweight fusion dataset of multi-modal time-series data through knowledge distillation, collect raw data from different modalities, unify the format and label time stamps, remove noise and fill missing values, normalize numerical data, and encode categorical data;

[0064] Compress it into a lightweight tensor form through knowledge distillation, complete time stamp alignment, and combine the compressed features with the original data labels to form a structured lightweight fusion dataset of multi-modal time-series data;

[0065] Assume that knowledge distillation is divided into a teacher model and a student model. The teacher model conducts end-to-end training on the original multi-source heterogeneous data to learn the potential associations and temporal features in the data. The student model learns the output and intermediate features of the teacher model through distillation while compressing the data into a tensor form.

[0066] As Figure 2 shown, the steps of the time-domain causal chain analysis module for mining the dynamic causal relationships between variables include: obtaining the multi-modal temporal data of the cross-modal distillation data fusion module;

[0067] Specifically, assume that the temporal data of each modality M is , the m-modal index is denoted as , assume the causal function , which measures the causal strength of variable on . The larger the value, the stronger the causal relationship. Calculate the dynamic causal weight of variables i and j at time t. The implementation formula of the dynamic causal weight is:

[0068] ,

[0069] In the formula, represents the dynamic causal weight of variables i and j at time t, T represents the time window length, represents the time window index, represents the time decay factor, m represents the modal index, M represents the total number of modalities, represents the coupling coefficient of modality m, , which reflects the contribution weight of the m-th modality to the causal relationship, represents the causal function, which measures the causal strength of variable on .

[0070] Specifically, the steps for dynamically updating the causal strength of the time-domain causal chain analysis module include: the dynamic causal weight is the causal strength at the current time step;

[0071] According to , calculate the dynamic causal strength from variable i to j. Assume the historical causal strength is . Combining the current causal weight and the historical causal strength , through the smoothing coefficient , balance the influence of the current weight and the historical causal strength. The dynamic update of the causal strength is implemented as:

[0072] ,

[0073] In the formula, Denotes the dynamic causal strength, the dynamic causal strength from variable i to j, Denotes the smoothing coefficient, Denotes the non-linear activation function, denoted as ;

[0074] Compresses the causal strength into the range (0, 1) as above, simulating the non-linear response of the biological system to causal relationships;

[0075] Specifically, the final causal strength Based on the weighted average of the current causal weight and the historical causal strength, recursive correction is used to avoid misjudgment of causal relationships caused by short-term noise, Denotes the historical causal strength.

[0076] For the time-domain causal chain analysis module, let the causal graph be ;

[0077] According to the dynamic causal strength , set the dynamic threshold and the node importance , and dynamically adjust the existence of the edge through the dynamic threshold and the node importance ;

[0078] Specifically, the causal graph topology update is implemented as:

[0079] ,

[0080] In the formula, Denotes the causal graph at time t, Denotes the causal edge of variable , Denotes the dynamic causal strength, Denotes the dynamic threshold, Denotes the node importance, Denotes the maximum importance value of all nodes at the current moment, high-importance nodes are close to , and the threshold of the causal edge decreases, and the dynamic threshold is dynamically adjusted according to the causal strength distribution.

[0081] The beneficial effects of the above technical solutions are: by weighted-averaging the current causal weight and the historical causal strength, and combining the non-linear activation function to compress the causal strength into the range (0, 1), short-term noise is effectively suppressed and the model stability is improved;

[0082] Dynamically adjust the contribution weights of different modalities to avoid biases from a single data source;

[0083] Dynamic causal weights display the causal paths between variables in the form of a time series, helping doctors intuitively understand the disease progression mechanism;

[0084] The core driving factors are screened out through recursive correction, providing a basis for personalized intervention.

[0085] The time-domain causal chain analysis module provides for the risk prediction module through the multimodal coupling coefficient and the node importance adjustment mechanism:

[0086] At the individual level: the personalized risk factor weights of patients; at the population level: the dynamic transmission paths of regional epidemics. The causal chain analysis module corrects the errors caused by outdated data or noise interference in the multi-level disease risk prediction module by dynamically updating the causal map.

[0087] For example, in practice, in diabetes detection, the dynamic causal chain can analyze the causal relationship among "dietary intake, exercise frequency, and blood sugar level" in real time, dynamically adjust the insulin dose recommendation, and reduce the risk of complications.

[0088] In the multi-level disease risk prediction module, the steps for individual disease risk scoring include: according to the multimodal time-series data of the cross-modal distillation data fusion module and the causal map of the time-series causal chain analysis module the causal intensity in the feature importance, combined with the dynamic weight determining the individual i's influence on the variable j ;

[0089] Specifically, the dynamic risk score of individual i is implemented as:

[0090] ,

[0091] In the formula, represents the individual dynamic risk score, the dynamic risk score of individual i at time t, d represents the total number of variables, represents the latest update time of variable j, represents the time decay factor, represents the dynamic weight of individual i on variable j, is represented as ;

[0092] Among them, , represents the hyperparameter, balancing the contributions of causal intensity and feature importance, represents calculating the feature importance through the multimodal time-series data , represents quantifying the current risk contribution of variable j.

[0093] In this embodiment, the steps for population risk scoring include: setting the population characteristics as , according to the individual dynamic risk score , combined with the group characteristics for , calculate the risk score of group p, implemented as:

[0094] ,

[0095] wherein, represents the group risk score, the risk score of group p at time t, represents the number of individuals in group p, , represents the mean and standard deviation of group p on modality m, represents the spatial smoothing function, is represented as , suppressing the influence of outliers on the group score, represents the group characteristic weighting factor.

[0096] According to the group risk score , adjust the weight of the individual dynamic risk score through the feedback mechanism ;

[0097] Specifically, the risk score feedback optimization is implemented as:

[0098] ,

[0099] In the formula, represents the updated weight of individual i for variable j at time t + 1, represents the learning rate, represents the dynamic weight of individual i for variable j, represents the adjustment increment of the weight of individual i for variable j at time t, is represented as , represents the feedback intensity coefficient, represents the gradient of the group risk score with respect to the individual weight.

[0100] The multi - level disease risk prediction module utilizes the time - decay factor and smoothing coefficient in the causal graph to improve the stability and accuracy of prediction.

[0101] The beneficial effects of the above - mentioned technical solution are as follows: The individual dynamic risks obtained by the multi - level disease risk prediction module are integrated into the group risk data for a more perfect and stable group risk score, and the individual data is updated and corrected according to the group risk score, improving the accuracy.

[0102] The multi - level disease risk prediction module is adjusted according to the dynamic causal strength in the causal map. When the causal strength of a certain disease in the causal map increases, the weight of this variable in the individual model is significantly higher than other variables. When the dynamic causal strength in the causal map is updated due to new data, the risk prediction module will automatically recalculate the individual and population scores;

[0103] Over time, the strength of the causal relationship may gradually increase or decrease, and the risk prediction module maintains the timeliness of prediction through dynamic weight adjustment.

[0104] This application responds to data changes in real - time through a time - decay factor and a feedback mechanism to avoid lag;

[0105] This application improves robustness by suppressing group outliers and using a spatial smoothing function to reduce the interference of accidental errors;

[0106] This application has an individual - group closed - loop feedback. For example, after high - risk groups trigger policy interventions, the weights of the individual model are corrected in reverse, and the coordination of precision medicine and public health strategies significantly improves the accuracy of risk prediction and the intervention efficiency;

[0107] Among them, the disease critical state real - time monitoring module collects vital sign data in real - time through monitoring devices and inputs the real - time collected vital sign data into the multi - level disease risk prediction module for real - time risk scoring;

[0108] Calculate a dynamic threshold based on the patient's historical vital sign data and risk score, and dynamically adjust the threshold according to the real - time risk score distribution. According to the degree of deviation of the real - time individual dynamic risk score from the threshold, divide the warning level. When the real - time individual dynamic risk score exceeds the dynamic threshold, it is determined that the patient has entered the critical state.

[0109] The adaptive intervention strategy generation module obtains the risk score of the multi - level disease risk prediction module and the critical state data of the disease critical state real - time monitoring module, and integrates the risk score, critical state data, and patient characteristics into a unified strategy input vector;

[0110] Define basic intervention rules according to clinical guidelines and expert experience, adapt the rules in combination with the patient's individual characteristics, match the rule base according to the current state, generate candidate intervention plans, and model the intervention strategy generation problem as a Markov decision process;

[0111] Train a policy network through a deep reinforcement learning algorithm, and dynamically update the reward function weights. Input the current state into the trained policy network to output the optimal intervention action.

[0112] Working principle: By integrating multi-source heterogeneous data, unifying the format and annotating timestamps, through preprocessing steps such as denoising, missing value filling, and normalization, the data is compressed into a lightweight tensor form, the high-dimensional complex data is compressed into a low-dimensional feature tensor, the timestamps of different modalities are aligned, and a structured and standardized multi-modal time series dataset is generated;

[0113] Based on the fused multi-modal time series data, the time-domain causal chain analysis module mines the dynamic causal relationships between variables, constructs and dynamically updates the causal graph, quantifies the causal strength between variables through the time decay factor and the smoothing coefficient, adjusts the weight and existence of the graph edges according to the dynamic causal strength, and the causal threshold of high-importance nodes is lower, enhancing the sensitivity to the critical path. The multi-level disease risk prediction module combines the multi-modal data features and the causal strength of the causal graph, calculates the dynamic weight of an individual for the critical variables, generates an individual risk score, based on the individual score and population characteristics, suppresses the influence of outliers through the spatial smoothing algorithm, outputs the disease trend prediction at the population level, and adjusts the individual weight in reverse through the population score to achieve continuous iteration of the prediction model; Real-time vital sign data is obtained through wearable devices or hospital monitoring systems and input into the risk prediction module to generate a real-time risk score. According to the patient's historical data and the current risk score distribution, the warning threshold is dynamically corrected to avoid false alarms / missing alarms caused by a fixed threshold. The warning level is divided according to the degree of deviation of the risk score from the threshold. When the score exceeds the threshold, it is determined that the patient enters a critical state. The adaptive intervention strategy generation integrates the risk score, critical state data, and patient characteristics into a unified input vector, constructs a basic intervention rule library based on clinical guidelines, adapts the rules according to the patient's individual differences, models the intervention strategy as a Markov decision process, trains the policy network through a deep reinforcement learning algorithm, dynamically adjusts the weight of the reward function, and outputs the optimal intervention action. Summarize the feedback from doctors and patients on the intervention plan, including subjective evaluations and objective data, and use the feedback data to iteratively update the multi-modal time series data of the cross-modal data fusion module through an encrypted federated learning framework to improve the data quality and prediction accuracy.

[0114] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

[0115] The above description of the present invention and its embodiments is not restrictive. What is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In short, if those of ordinary skill in the art are inspired by it and design similar structural methods and embodiments without creative efforts without departing from the purpose of the present invention, they should all fall within the protection scope of the present invention.

Claims

1. An analysis and prediction system based on disease data, characterized in that: It includes a cross-modal distillation data fusion module, a time-domain causal chain analysis module, a multi-level disease risk prediction module, a real-time monitoring module for disease critical states, an adaptive intervention strategy generation module, and a clinical collaboration and feedback optimization module; The time-domain causal chain analysis module is used to obtain the dynamic causal relationships among variables by mining multi-modal time-series data, and construct and update the causal map; The multi-level disease risk prediction module is used to combine the multi-modal time-series data of the cross-modal distillation data fusion module and the causal map of the time-domain causal chain analysis module to hierarchically predict the disease risk scores of individuals and groups; The real-time monitoring module for disease critical states is used to obtain real-time vital sign data and the risk scores of the multi-level disease risk prediction module, and dynamically adjust the critical threshold to trigger an alarm.

2. The analysis and prediction system based on disease data according to claim 1, wherein: The steps of the time-domain causal chain analysis module for mining the dynamic causal relationship between variables include: obtaining the multi-modal time-series data of the cross-modal distillation data fusion module, and setting the time-series data of each modality M as , and the m-modal index is expressed as . Let the causal function be . Calculate the dynamic causal weight of variables i and j at time t. The implementation formula of the dynamic causal weight is: ， In the formula, represents the dynamic causal weight of variables i and j at time t, T represents the time window length, represents the time window index, represents the time decay factor, m represents the modality index, and M represents the total number of modalities, represents the coupling coefficient of modality m, represents the causal function that measures the variable to causal strength.

3. The analysis and prediction system based on disease data according to claim 2, wherein: The causal intensity dynamic update step of the time-domain causal chain analysis module includes: dynamic causal weights is the causal intensity at the current time step. According to the dynamic causal weights calculate the dynamic causal intensity from variable i to j , let the historical causal intensity be , combined with the current dynamic causal weights and the historical causal intensity , the dynamic update of the causal intensity is implemented as: ， In the formula, represents the dynamic causal strength, the dynamic causal strength from variable i to j, represents the smoothing coefficient, represents the non-linear activation function, denoted as , represents the historical causal strength.

4. The analysis and prediction system based on disease data according to claim 3, wherein: The time-domain causal chain analysis module sets the causal graph as , and the causal graph topology update steps include: According to the dynamic causal strength , set the dynamic threshold and the node importance , through the dynamic threshold and the node importance dynamically adjust the existence of edges, and the causal graph topology update is implemented as: ， In the formula, represents the causal graph, the causal graph at time t, represents the variable of the causal edge, represents the dynamic causal strength, represents the dynamic threshold, represents the node importance, represents the maximum importance value of all nodes at the current moment.

5. The analysis and prediction system based on disease data according to claim 1, wherein: For the multi-level disease risk prediction module, the steps for individual disease risk scoring include: based on the multi-modal time-series data of the cross-modal distillation data fusion module and the causal graph of the time-series causal chain analysis module the feature importance of the dynamic causal strength in it, combined with the dynamic weight that determines individual i's influence on variable j , the dynamic risk score of individual i is implemented as: ， In the formula, represents the individual dynamic risk score, the dynamic risk score of individual i at time t, d represents the total number of variables, represents the latest update time of variable j, represents the time decay factor, represents the dynamic weight of individual i for variable j, represents the current risk contribution of quantifying variable j.

6. The analysis and prediction system based on disease data according to claim 5, characterized in that: For the multi-level disease risk prediction module, the steps for group risk scoring include: setting the group characteristics as , based on the individual dynamic risk score , combining the group characteristics as , calculating the risk score of group p, which is implemented as: ， In the formula, represents the population risk score, the risk score of population p at time t, represents the number of individuals in population p, , represents the mean and standard deviation of population p on modality m, represents the spatial smoothing function, represents the population characteristic weighting factor.

7. The analysis and prediction system based on disease data according to claim 6, wherein: For the multi-level disease risk prediction module, the risk score feedback optimization steps are as follows: Based on the population risk score , adjust the weights of the individual dynamic risk scores through a feedback mechanism , and the risk score feedback optimization is implemented as: ， In the formula, represents the updated weight of individual i for variable j at time t + 1, represents the learning rate, represents the dynamic weight of individual i for variable j, represents the adjustment increment of the weight of individual i for variable j at time t.

Images