Construction and application of a dynamic causal model for chronic disease management

Abstract

The present application proposes a construction and application method of a dynamic causal model for chronic disease management, which continuously collects multimodal dynamic health data of patients, takes multimodal dynamic health data as a node set, takes the causal relationship between multimodal dynamic health data as an edge set, takes the causal strength between multimodal dynamic health data as a weight matrix, and constructs a dynamic causal model based on the node set, edge set and weight matrix; for new multimodal dynamic health data, calculates the marginal causal effect value of the current new multimodal dynamic health data with all the node sets in the dynamic causal model, and updates the dynamic causal model based on the marginal causal effect value; for each time step, calculates the comprehensive causal weight of each side of each time step based on the long-term component characteristics and short-term component characteristics of the multimodal dynamic health data corresponding to the same side, updates the dynamic causal model based on the comprehensive causal weight, and proposes an application method for the dynamic causal model.

Description

Technical Field

[0001] The present application relates to the field of chronic disease management, and in particular to a method for constructing and applying a dynamic causal model for chronic disease management. Background Art

[0002] In the field of chronic disease management, accurately understanding and grasping the causal relationship between health variables is crucial to formulating effective intervention strategies and improving management outcomes.

[0003] However, the causal management network established by existing technologies for causal relationships has many shortcomings in this regard:

[0004] Traditional static causal networks, such as Bayesian networks and structural equation models, establish causal relationships based on fixed time points or static data. When the health status of patients with chronic diseases changes over time, this static model cannot promptly reflect the impact of these changes on causal relationships. In addition, the existing technology only establishes simple regression models based on overall correlation or directly relies on historical data, which ignores the problem of short-term and long-term factors for independent modeling. Such methods cannot effectively distinguish the causal characteristics of short-term fluctuations (such as postprandial blood sugar fluctuations) and long-term trends (such as changes in glycated hemoglobin levels). For example, postprandial blood sugar fluctuations may be affected by short-term factors such as immediate diet and exercise, while glycated hemoglobin levels reflect long-term blood sugar control. It is difficult for a single model to accurately characterize the causal relationship between the two and their different impacts at the same time.

[0005] Moreover, there are many deficiencies in the application of causal networks:

[0006] On the one hand, traditional causal inference technology usually does not dynamically generate causal chains in combination with specific scenarios, but outputs a global causal network. This makes it difficult to meet the personalized intervention needs in chronic disease management, because in actual clinical applications, the causal relationships that need to be focused on are different for different patients and different stages of the disease. The lack of causal chain generation technology for specific target variables and causal variables has led to poor operability of existing causal analysis results in clinical practice, and it is difficult for clinicians to quickly obtain key information from complex causal networks to develop personalized intervention plans.

[0007] In addition, existing intervention optimization methods mainly rely on direct correlation analysis, which makes it difficult to comprehensively evaluate the direct and indirect effects of intervention variables. In actual chronic disease management, an intervention variable may indirectly affect the target variable through multiple pathways, and only considering direct correlation will lead to inaccurate evaluation of the intervention effect. At the same time, there is a lack of systematic methods to combine intervention effects with intervention costs to optimize the overall cost-effectiveness of intervention paths. For example, when choosing drug treatment or lifestyle intervention, it is difficult to formulate a scientific, reasonable and cost-effective intervention strategy without comprehensive consideration of intervention costs (such as drug prices, time costs, difficulty of behavioral adjustment, etc.) and possible effects. Moreover, after the implementation of the intervention, there is a lack of real-time feedback mechanism, and the intervention path and causal weight cannot be dynamically adjusted according to the actual effect, making it difficult to continuously optimize the intervention plan to adapt to the dynamic changes of patients. Summary of the invention

[0008] The present application provides a method for constructing and applying a dynamic causal model for chronic disease management, aiming to solve the problems of causal network selection in the prior art and propose a dynamic causal model suitable for chronic disease management.

[0009] To achieve the above objectives, in the first aspect, this solution provides a method for constructing a dynamic causal model for chronic disease management, comprising the following steps:

[0010] Continuously collect multimodal dynamic health data of patients, and decompose the multimodal dynamic health data into long-term component features and short-term component features according to time characteristics;

[0011] Taking multimodal dynamic health data as a node set, the causal relationship between multimodal dynamic health data as an edge set, and the causal strength between multimodal dynamic health data as a weight matrix, a dynamic causal model is constructed based on the node set, edge set and weight matrix;

[0012] For the new multimodal dynamic health data, the marginal causal effect values ​​of the current new multimodal dynamic health data and all the node sets in the dynamic causal model are calculated, and the dynamic causal model is updated based on the marginal causal effect values;

[0013] For each time step, the comprehensive causal weight of each side of each time step is calculated according to the long-term component characteristics and short-term component characteristics of the multimodal dynamic health data corresponding to the same side, and the dynamic causal model is updated based on the comprehensive causal weight.

[0014] Secondly, this solution provides an application method of a dynamic causal model for chronic disease management, including the following steps:

[0015] Select the cause node and target node;

[0016] Searching for potential edges from the cause node to the target node as potential dynamic causal chains, where the causal strength of the potential edges is greater than a set threshold;

[0017] The potential dynamic causal chain with the greatest causal strength is selected as the dynamic causal chain.

[0018] In a third aspect, the present application provides an application method of a dynamic causal model for chronic disease management, comprising the following steps:

[0019] Select at least two intervention nodes and a target node;

[0020] Searching for all influence edges from each intervention node to the target node, wherein the influence edges include direct edges from each intervention node to the target node, and indirect edges from the intervention node to the target node through other nodes;

[0021] Calculate the influence of each intervention node based on all influencing edges and intervention changes of the intervention node;

[0022] The intervention cost-effectiveness of each intervention node is calculated based on the influence of each intervention node and the intervention cost of the corresponding intervention node;

[0023] Select intervention nodes with high cost-effectiveness.

[0024] Different from the prior art, the main contributions and innovations of the present invention are as follows:

[0025] Strong dynamic adaptability: The present invention can capture the dynamic changes of causal relationships of multimodal dynamic health data in chronic disease management in real time, and has the ability to incrementally update nodes, edges and weights, avoiding the high computational cost of traditional causal network reconstruction, and significantly improving the adaptability of the model to new variables, scene changes and data updates. Accurate characterization of short-term and long-term impacts: Through hierarchical modeling of short-term component features and long-term component features, it is possible to distinguish the different effects of short-term fluctuations and long-term trends in multimodal dynamic health data, and accurately characterize the causal relationship between immediate behavior (such as postprandial blood sugar fluctuations) and long-term trends (such as changes in glycated hemoglobin). This multi-time scale modeling greatly improves the inference accuracy and clinical interpretability of the causal network. Improve the pertinence and reliability of causal chain generation: The present invention dynamically extracts the dynamic causal chain of the target variable in a specific scenario, and generates a more reliable and pertinent dynamic causal chain through strict causal weight screening and path scoring mechanism, which can help clinicians quickly lock in key influencing factors and provide efficient support for personalized decision-making in chronic disease management. Optimize the selection of intervention paths: The present invention combines intervention effects with intervention costs, optimizes the cost-effectiveness of intervention paths through a comprehensive evaluation of direct and indirect influences, and the real-time feedback mechanism enables the intervention paths to be dynamically adjusted according to actual effects, ensuring continuous optimization of the program and stability of implementation effects.

[0026] Details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0028] Figure 1 This is a logical diagram of the method for constructing a dynamic causal model for chronic disease management provided in this solution.

[0029] Figure 2 It is a logical diagram of the method for constructing a dynamic causal model for chronic disease management according to Example 2 of the present application.

[0030] Figure 3 It is a logical diagram of the method for constructing a dynamic causal model for chronic disease management according to Example 3 of the present application. DETAILED DESCRIPTION

[0031] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with one or more embodiments of this specification. Instead, they are merely examples of devices and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0032] It should be noted that: in other embodiments, the steps of the corresponding method are not necessarily performed in the order shown and described in this specification. In some other embodiments, the steps included in the method may be more or less than those described in this specification. In addition, a single step described in this specification may be decomposed into multiple steps for description in other embodiments; and multiple steps described in this specification may be combined into a single step for description in other embodiments.

[0033] Embodiment 1

[0034] like Figure 1 As shown, the embodiment of the present application provides a method for constructing a dynamic causal model for chronic disease management, comprising the following steps:

[0035] Continuously collect multimodal dynamic health data of patients, and decompose the multimodal dynamic health data into long-term component features and short-term component features according to time characteristics;

[0036] Taking multimodal dynamic health data as a node set, the causal relationship between multimodal dynamic health data as an edge set, and the causal strength between multimodal dynamic health data as a weight matrix, a dynamic causal model is constructed based on the node set, edge set and weight matrix;

[0037] For the new multimodal dynamic health data, the marginal causal effect values ​​of the current new multimodal dynamic health data and all the node sets in the dynamic causal model are calculated, and the dynamic causal model is updated based on the marginal causal effect values;

[0038] For each time step, the comprehensive causal weight of each side of each time step is calculated according to the long-term component characteristics and short-term component characteristics of the multimodal dynamic health data corresponding to the same side, and the dynamic causal model is updated based on the comprehensive causal weight.

[0039] The dynamic causal model constructed by this solution is specifically designed for chronic disease management settings. It can be dynamically updated according to the changes in multimodal dynamic health data collected in real time, thereby effectively capturing the dynamic evolution of causal relationships over time. Especially for scenarios where multimodal dynamic health data is added or removed, the present invention adopts an incremental update strategy to avoid full graph reconstruction, greatly improving computing efficiency and network adaptability. This dynamic network provides technical support for real-time inference and adjustment in complex chronic disease health management scenarios.

[0040] In the step of "continuously collecting multimodal dynamic health data of patients", the multimodal dynamic health data includes multimodal health data representing the characteristics of chronic diseases for the same patient. The subsequent dynamic causal model can analyze the causal relationship between different multimodal dynamic health data and chronic diseases based on the multimodal dynamic data.

[0041] In some embodiments, multimodal dynamic health data includes one or any combination of physiological data, behavioral data, and environmental data. Physiological data is data that characterizes the patient's physiological characteristics, such as continuous indicators such as blood sugar, heart rate, and blood pressure; behavioral data is data that characterizes the patient's behavioral characteristics, such as discrete records of diet logs, exercise volume, sleep duration, etc.; environmental data is data that characterizes the environment in which the patient lives, such as temperature, air quality, etc.

[0042] In some embodiments, multimodal dynamic health data from different sources are time-aligned and organized into time series data according to the time sequence. Specifically, this solution aligns multimodal dynamic health data from different sources according to the set target time step, and interpolates and supplements the missing values. The formula for interpolation and supplementation is as follows: ;

[0043] in is the multimodal dynamic health data x corresponding to the target time step t i The interpolation result, t prev is the closest observation time point before the target time step t, t next is the closest observation time point after the target time step t, is the multimodal dynamic health data x i At time point t next The observed value on is the multimodal dynamic health data x i At time point t prev The observed value on .

[0044] The purpose of time alignment is to align multimodal dynamic health data from different sources and different sampling frequencies to the same time point. For example, heart rate data is sampled every second, blood glucose data is sampled every 5 minutes, and exercise data is recorded every 10 minutes. So when a unified time step is set, such as 1 minute as the unit of time alignment, then for heart rate data, take 1 value every 1 minute (such as the average value of 1 minute), for blood glucose data, 4 values ​​are filled between each data. For exercise data, 9 values ​​are filled between each data.

[0045] The present invention separates the short-term and long-term component characteristics of multimodal dynamic health data and introduces multi-time-scale causal modeling. It can accurately characterize the immediate impact of short-term behaviors (such as diet and exercise) on health variables, and the cumulative effect of long-term factors (such as lifestyle and underlying chronic disease status) on health trends. This hierarchical modeling method significantly improves the accuracy and clinical interpretability of causal weights, and can meet the dual needs of short-term intervention and long-term monitoring in chronic disease management.

[0046] Correspondingly, in the step of "decomposing the multimodal dynamic health data into long-term component features and short-term component features according to the time characteristics", a sliding window is set according to the time characteristics of the multimodal dynamic health data, and the multimodal dynamic health data is decomposed into long-term component features and short-term component features based on the sliding window. The short-term component features of this solution are used to characterize the rapid changes in the multimodal dynamic health data, and the time scale is often minutes or hours; the long-term component features are used to characterize the overall change characteristics of the multimodal dynamic health data, and the time scale is often days, weeks, and quarters. For example, blood sugar is significantly affected by high frequency and short-term behaviors (such as diet and exercise), so the change in blood sugar within 1-2 hours after a meal is a short-term component feature; and the daily trend or long-term change of blood sugar is a long-term component feature, which is used to reflect the long-term control of chronic diseases.

[0047] Specifically, the formula for decomposing multimodal dynamic health data into long-term component features and short-term component features based on a sliding window is as follows:

[0048]

[0049] ;

[0050] in is the short-term component characteristic, is the long-term component feature, t is the time step, x t is the variable corresponding to time t, It is a sliding window, and the sliding window length can be flexibly set according to different clinical scenarios.

[0051] It should be noted that this scheme artificially distinguishes short-term component features from long-term component features based on the clinical significance, time characteristics and medical experience of different multimodal dynamic health data. For example, for diabetic patients, the rapid change of blood sugar after a meal is a short-term component feature, and the change of glycated hemoglobin is a long-term component. Therefore, a 2-hour time window can be selected to capture the fluctuation of blood sugar after a meal as a short-term component feature, and a 1-week window can be selected to capture the change trend of glycated hemoglobin as a long-term component feature. Moreover, the length of the time window set by this scheme can also be optimized and adjusted according to subsequent feedback. For example, if the actual feedback data shows that the time window of the short-term component feature of 2 hours is not enough to capture rapid changes, the set time window can be adjusted to 1 hour. If the actual feedback data shows that the time window of the long-term component feature is too large, resulting in the signal being too smooth and no trend can be seen, the length of the time window can be appropriately reduced.

[0052] After preprocessing the collected multimodal dynamic health data, this solution can build a dynamic causal model based on the multimodal dynamic health data. The dynamic causal model constructed by this solution is expressed as G(V,E,W), where the node set V represents the variables represented by the multimodal dynamic health data, the edge set E represents the causal relationship between the variables, and W represents the causal weight of the edge in the edge set, which corresponds to the weight matrix. Specifically, the node set V={x1,x2,…,x n}, representing the variables extracted from the multimodal dynamic health data (such as blood sugar, diet, exercise); the edge set E ⊆ V×V, represents the causal relationship between the variables. The variables with potential causal relationships can be screened as candidate edges through the edge set E. Each edge e ij ∈E represents the variable x i For variable x j The causal relationship of , w ij ∈W represents edge e ij causal strength.

[0053] In the step of "taking the causal relationship between multimodal dynamic health data as the edge set", the multimodal dynamic health data is subjected to conditional independence verification to remove irrelevant edges and obtain candidate edges. For the correlation between the two multimodal dynamic health data corresponding to all candidate edges, the candidate edges whose correlation is greater than the set threshold are retained as the edges of the edge set.

[0054] in represents the ith multimodal dynamic health data, corresponding to the ith variable, represents the j-th multimodal dynamic health data, corresponding to the j-th variable, represents the covariance between two variables, represents the variance of the ith variable, represents the variance of the jth variable, is the correlation between two variables;

[0055] The correlations corresponding to all candidate edges form the following correlation matrix:

[0056] .

[0057] In some embodiments, in order to enhance the medical nature of the dynamic causal model, prior edges are supplemented in the edge set based on medical field knowledge, and the supplementation methods include but are not limited to manual labeling: experts directly define key causal relationships; rule engine: based on existing medical-related rule engines, unlabeled causal edges are automatically supplemented; knowledge base: similar to the rule engine, the causal edges are automatically supplemented using a medical knowledge base that stores structured medical causal relationships.

[0058] In the step of "taking the causal strength between multimodal dynamic health data as the weight matrix", the correlation between the multimodal dynamic health data is used as the initial weight, and the quotient of the initial weight and the variance of the current multimodal dynamic health data is used as the marginal causal effect value between the multimodal dynamic health data. The initial weight is revised based on the marginal causal effect value to obtain the causal strength between the multimodal dynamic health data.

[0059] Specifically, the calculation formula is as follows:

[0060]

[0061] in is the correlation between two variables, is the marginal causal effect value between two variables, represents the variance of the ith variable.

[0062] ;

[0063] in is the causal strength between multimodal dynamic health data, η is the balance factor between correlation and marginal causal effect, and its value range is [0,1].

[0064] In some embodiments, a professional toolkit (here, networkx) may be used to construct a dynamic causal model.

[0065] It should be noted that, unlike traditional static causal networks (such as Bayesian networks and structural equation models), which usually construct causal relationships based on fixed time points or static data, which makes it impossible to effectively cope with the changes in causal relationships over time in chronic disease management, and it is difficult to deal with the time lag effect between health variables and the dynamic adjustment of causal weights, the dynamic causal model proposed in this scheme realizes dynamic adjustment in chronic disease management. The update and optimization of the dynamic causal model of this scheme is aimed at coping with the complex needs brought about by the changes in causal relationships over time, changes in scenarios, new data input, and adjustments to network structures, while improving the accuracy and efficiency of inference. In the complex system of chronic disease management, changes in nodes, edges, and weights require the causal network to have the ability to adapt in real time and continuously optimize. In order to achieve efficient and accurate network adjustment, it can be updated at three levels: node set V, edge set E, and causal weight matrix W.

[0066] When new multimodal dynamic health data is introduced, the dynamic causal model can be dynamically expanded based on the relationship between the new multimodal dynamic health data and the existing node set; and when some multimodal dynamic health data loses its influence or is confirmed to be irrelevant, the irrelevant edges of the dynamic causal model can be cleared.

[0067] Specifically, in the step of "for the new multimodal dynamic health data, calculating the marginal causal effect values ​​of the current new multimodal dynamic health data and all the node sets in the dynamic causal model, and updating the dynamic causal model based on the marginal causal effect values",

[0068] For each new multimodal dynamic health data, the marginal causal effect value of the current new multimodal dynamic health data and all the node sets in the dynamic causal model is calculated. The formula is as follows: ;

[0069] The new multimodal dynamic health data is represented as the i-th variable , Represents the current new multimodal dynamic health data For the kth variable The marginal causal effect value of express and The correlation between represents the variance of the ith variable, Represents all nodes in the node set.

[0070] When the marginal causal effect value is greater than a certain threshold, an edge is established between the new multimodal dynamic health data and the corresponding node, and the correlation between the new multimodal dynamic health data and the associated node is used as the initial weight of the edge:

[0071] .

[0072] In the step of "for each time step, calculate the comprehensive causal weight of each side at each time step based on the long-term component characteristics and short-term component characteristics of the multimodal dynamic health data corresponding to the same side, and update the dynamic causal model based on the comprehensive causal weight", the short-term causal weights of the short-term component characteristics of the two multimodal dynamic health data corresponding to the same side are calculated, the long-term causal weights are calculated based on the time lag window and the long-term component characteristics of the two multimodal dynamic health data corresponding to the same side, and the comprehensive causal weight of the current time step is calculated based on the short-term causal weights, the long-term causal weights and the comprehensive causal weight of the previous time step on the current side.

[0073] Specifically, the formula for calculating the short-term causal weights of the short-term component characteristics of two multimodal dynamic health corresponding to the same side is as follows:

[0074] ;

[0075] in is the short-term component characteristic of the ith variable, is the long-term component characteristic of the j-th variable, represents the correlation between the short-term component characteristics of the i-th variable and the long-term component characteristics of the j-th variable, is the variance of the short-term component characteristics of the ith variable, It is the short-term causal weight between the i-th variable and the j-th variable.

[0076] The formula for calculating the long-term causal weight based on the time lag window and the long-term component characteristics of the two multimodal dynamic health data corresponding to the same side is as follows: ;

[0077] ;

[0078] in is the time lag window, represents the long-term component characteristics of the i-th variable after lagged by k time periods, is the long-term component characteristic of the j-th variable, is the correlation between the long-term component characteristics of the i-th variable after the k-th time lag and the long-term component characteristics of the j-th variable, is the variance of the long-term component characteristics of the ith variable after lagged k times, is the hysteresis weight, where λ is the decay rate, which controls the decay amplitude of the weight. Represents the long-term causal weight between the i-th variable and the j-th variable.

[0079] The time lag window of this scheme is used to capture the time range of the lag effect, that is, to examine the variable x i In the past lag The potential impact on the variable xj within a time step is generally set directly by the advice of experts based on relevant clinical and field experience. For example, according to experience, the delayed effect of diet on blood sugar may reach its peak within 2 hours, so setting T lag When the actual data comes in, the performance of the system can be evaluated in actual applications based on the hysteresis window range for further verification and adjustment.

[0080] The formula for calculating the comprehensive causal weight of the current time step based on the short-term causal weight, the long-term causal weight and the comprehensive causal weight of the current edge at one time step is as follows: ;

[0081] Among them, γ is the balance factor between the short-term causal weight and the long-term causal weight, and β is the influencing factor of the comprehensive causal weight of the previous time step. The comprehensive causal weight update of this scheme at the t+1 time step takes into account the short-term causal weight, the long-term causal weight, and the comprehensive causal weight of the previous time step.

[0082] In the step of "updating the dynamic causal model based on the comprehensive causal weight", the total causal weight of the current node is calculated based on the comprehensive causal weights of the different edges connecting each node. If the total causal weight is less than the set threshold, the current node and the corresponding connected edges are removed. This solution removes variables with no observation data for a long time and irrelevant variables in this way to optimize the network architecture of the dynamic causal model.

[0083] Specifically, the total causal weight of each node is as follows:

[0084] ;

[0085] in represents the comprehensive causal weight of the edge between the kth node and the i-th node, represents the comprehensive causal weight of the edge between the kth node and the jth node, is the total causal weight of the current node.

[0086] Embodiment 2

[0087] The first embodiment of the present solution constructs a dynamic causal model for chronic disease management that can be changed dynamically. In the context of chronic disease management, it is necessary to further generate causal chains related to specific scenarios based on the dynamic causal model to better support decision-making and personalized intervention. In other words, the present solution can extract the most critical causal path for the target variable from the complex causal network based on the constructed dynamic causal model, and dynamically adjust the structure of the causal chain and the intervention strategy in a specific scenario.

[0088] Specifically, Figure 2 As shown, this solution provides an application method of a dynamic causal model for chronic disease management, which is implemented based on the dynamic causal model constructed in Example 1 and includes the following steps:

[0089] Select the cause node and target node;

[0090] Searching for potential edges from the cause node to the target node as potential dynamic causal chains, where the causal strength of the potential edges is greater than a set threshold;

[0091] The potential dynamic causal chain with the greatest causal strength is selected as the dynamic causal chain.

[0092] This solution is based on a dynamic causal model to extract the target causal chain from the cause node to the target node in a specific chronic disease management scenario. The dynamic causal chain refers to a causal path composed of key variables generated based on the dynamic causal model in a specific scenario, which represents the dynamic evolution relationship from the cause (such as living habits) to the result (such as disease status). This dynamic causal chain can help identify key influencing factors, provide decision support, and optimize intervention strategies. In different scenarios of chronic disease management, each scenario will have specific cause nodes and target nodes that need to be paid attention to. For example, in diabetes management, it may be necessary to analyze the causal relationship between dietary habits (cause node) and blood sugar fluctuations (target node); in the intervention optimization of hypertension, the target node may be the blood pressure level, and the cause node may include behavioral data such as sodium intake and exercise. To extract a dynamic causal chain from a dynamic causal model, it is necessary to focus on the target node and select the most representative path based on the weight to reveal how the cause node affects the target node through intermediate factors.

[0093] Specifically, the formula for searching for potential edges from the cause node to the target node as potential dynamic causal chains is:

[0094] ;

[0095] ;

[0096] in is the cause node, is the target node, is a potential dynamic causal chain, To set the threshold, is the i-th node and the jth node The causal strength between them corresponds to the comprehensive causal weight.

[0097] In some embodiments, depth-first search (DFS) is used to generate all potential dynamic causal chains from the cause node to the target node. The edges of the potential dynamic causal chains must be strictly limited to a screened subset to ensure that the generated potential dynamic causal chains have higher reliability and interpretability.

[0098] In some embodiments, selecting the potential dynamic causal chain with the greatest causal strength as the dynamic causal chain refers to selecting the potential dynamic causal chain with the greatest sum of the causal strengths of all edges on the potential dynamic causal chain. The higher the score of the dynamic causal chain, the stronger the causal relationship on the path as a whole, the higher the explanatory power of the target node, and the higher the value of subsequent research and decision-making. For example, taking diabetes management as an example, the target node is the patient's fasting blood glucose level, and the cause node may be eating habits. Then the extraction of the dynamic causal chain can generate the following causal path:

[0099]

[0100] This pathway indicates that dietary habits ultimately affect fasting blood glucose levels through the regulation of weight changes and insulin sensitivity. Combined with the analysis of pathway weights, personalized recommendations for dietary adjustments can be further made.

[0101] As described above, Example 2 of the present scheme is based on a dynamic causal model, and dynamically extracts dynamic causal chains for specific target nodes and cause nodes, avoiding the redundancy and non-targetedness of traditional global causal analysis. By limiting the weight screening threshold and reliability verification of the edges in the dynamic causal chain, the dynamic causal chain generated by the present invention is more targeted and reliable, which can help clinicians quickly identify key influencing factors and formulate personalized intervention strategies.

[0102] Embodiment 3

[0103] Intervention optimization is one of the important applications of dynamic causal models in chronic disease management. It aims to develop the optimal intervention path and strategy by analyzing the direct and indirect effects of intervention variables on target variables. The core goal of intervention optimization is to select key intervention variables, determine the intervention range, and optimize the intervention path to maximize the state of the target variable while minimizing the intervention cost. In chronic disease management, intervention optimization can help develop personalized patient management plans, such as controlling blood sugar, lowering blood pressure, or improving lifestyle habits.

[0104] Correspondingly, such as Figure 3As shown, this solution provides an application method of a dynamic causal model for chronic disease management, which is implemented based on the dynamic causal model constructed in Example 1 and includes the following steps:

[0105] Select at least two intervention nodes and a target node;

[0106] Searching for all influence edges from each intervention node to the target node, wherein the influence edges include direct edges from each intervention node to the target node, and indirect edges from the intervention node to the target node through other nodes;

[0107] Calculate the influence of each intervention node based on all influencing edges and intervention changes of the intervention node;

[0108] The intervention cost-effectiveness of each intervention node is calculated based on the influence of each intervention node and the intervention cost of the corresponding intervention node;

[0109] Select intervention nodes with high cost-effectiveness.

[0110] In some embodiments, based on the finally selected intervention node as the causal node, the dynamic causal chain corresponding to the selected intervention node is obtained as the intervention causal chain according to the method of the second embodiment.

[0111] In the step of "searching for all influencing edges from each intervention node to the target node", if there is no direct edge between the current intervention node and the target node, a virtual edge is defined between the intervention node and the target node as a direct edge.

[0112] In the step of "calculating the influence of each intervention node based on all influencing edges and intervention changes of the intervention nodes", the direct influence of the current intervention node is calculated based on the direct edges, and the indirect influence of the current intervention node is calculated based on the indirect edges. The indirect influence and the direct influence are combined as the influence of the current intervention node.

[0113] Specifically, the formula for calculating the direct influence of the current intervention node based on the direct edge is as follows: ;

[0114] in For direct influence, is the causal strength between the i-th node (corresponding intervention node) and the target node, is the intervention change of the i-th node;

[0115] The formula for calculating the indirect influence of the current intervention node based on the indirect edge is as follows: ;

[0116] in For indirect influence, For all indirect edges from the intervention node to the target node, is the weighted product of the causal strengths of all indirect edges, represents the causal strength between the mth node and the nth node on the indirect edge, is the intervention change of the i-th node (corresponding intervention node).

[0117] The formula for combining indirect influence and direct influence as the influence of the current intervention node is as follows: ;

[0118] in is the influence of the current intervention node.

[0119] In the step of "calculating the intervention cost performance of each intervention node based on the influence of each intervention node and the intervention cost of the corresponding intervention change of the intervention node", the intervention cost of the intervention change of each intervention node is the product of the intervention change and the unit cost:

[0120] ;

[0121] in is the intervention cost corresponding to the intervention change of the intervention node, is the unit cost (such as drug price, time cost, difficulty of behavior adjustment, etc.), is the intervention change of the i-th node (corresponding intervention node).

[0122] This scheme uses the quotient of the influence of each intervention node and the intervention cost of the corresponding intervention node as the intervention cost performance of each intervention node. The calculation formula is as follows: ;

[0123] in is the intervention cost-effectiveness of each intervention node.

[0124] Of course, in some embodiments, after the intervention node is selected, the actual intervention effect can also be evaluated based on the change of the actual target node after the intervention is implemented, and the causal weight can be dynamically updated and adjusted based on the difference between the actual effect fed back and the predicted value.

[0125] The specific calculation formula is as follows: ;

[0126] ;

[0127] in is the actual change of the target node, is the value of the target node at time step t (corresponding to the value before the intervention), is the value of the target node at time step t+1 (corresponding to the value after intervention), is the causal strength between the current i-th node (intervention node) and the j-th node (corresponding to the target node), corresponding to the weight; is the adjustment factor.

[0128] As shown in Example 3 above, this solution proposes an intervention optimization method for a dynamic causal model, which comprehensively evaluates the direct and indirect effects of the intervention nodes, and calculates the intervention cost-effectiveness in combination with the intervention cost, to ensure the scientificity and economy of the intervention path. By real-time monitoring of the intervention effect, the present invention can dynamically adjust the causal weights and intervention plans, continuously optimize the intervention path, and improve the stability and efficiency of the intervention effect. For example, the combined path of dietary adjustment and exercise intervention can be dynamically optimized to achieve more accurate chronic disease management.

[0129] Those skilled in the art should understand that the technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0130] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for constructing a dynamic causal model for chronic disease management, characterized in that: The following steps are involved: Continuously collect multimodal dynamic health data of patients, and decompose the multimodal dynamic health data into long-term component features and short-term component features according to time characteristics; Taking multimodal dynamic health data as a node set, the causal relationship between multimodal dynamic health data as an edge set, and the causal strength between multimodal dynamic health data as a weight matrix, a dynamic causal model is constructed based on the node set, edge set and weight matrix; For the new multimodal dynamic health data, the marginal causal effect values ​​of the current new multimodal dynamic health data and all the node sets in the dynamic causal model are calculated, and the dynamic causal model is updated based on the marginal causal effect values; For each time step, the short-term causal weights of the short-term component features of the two multimodal dynamic healths corresponding to the same side are calculated, the long-term causal weights are calculated according to the time lag window and the long-term component features of the two multimodal dynamic healths corresponding to the same side, the comprehensive causal weight of the current time step is calculated based on the short-term causal weights, the long-term causal weights and the comprehensive causal weight of the previous time step of the current side, and the dynamic causal model is updated based on the comprehensive causal weights; The correlation between multimodal dynamic health data is used as the initial weight, the quotient of the initial weight and the variance of the current multimodal dynamic health data is used as the marginal causal effect value between the multimodal dynamic health data, and the initial weight is revised based on the marginal causal effect value to obtain the causal strength between the multimodal dynamic health data; The calculation formula is as follows: ； in is the correlation between two variables, is the marginal causal effect value between two variables, represents the variance of the i-th variable; ； in is the causal strength between multimodal dynamic health data, η is the balance factor between correlation and marginal causal effect, and its value range is [0,1]; The formula for calculating the short-term causal weights of the short-term component characteristics of two multimodal dynamic health corresponding to the same side is as follows: ； in is the short-term component characteristic of the ith variable, is the long-term component characteristic of the j-th variable, represents the correlation between the short-term component characteristics of the i-th variable and the long-term component characteristics of the j-th variable, is the variance of the short-term component characteristic of the ith variable, is the short-term causal weight between the i-th variable and the j-th variable; The formula for calculating the long-term causal weight based on the time lag window and the long-term component characteristics of the two multimodal dynamic health data corresponding to the same side is as follows: ; ； in is the time lag window, represents the long-term component characteristics of the i-th variable after lagged by k time periods, is the long-term component characteristic of the j-th variable, is the correlation between the long-term component characteristics of the i-th variable after the k-th time lag and the long-term component characteristics of the j-th variable, is the variance of the long-term component characteristics of the ith variable after lagged k times, is the hysteresis weight, where λ is the decay rate, which controls the decay amplitude of the weight. Represents the long-term causal weight between the i-th variable and the j-th variable; The formula for calculating the comprehensive causal weight of the current time step based on the short-term causal weight, the long-term causal weight and the comprehensive causal weight of the current edge at one time step is as follows: ; Where γ is the balance factor between the short-term causal weight and the long-term causal weight, and β is the influencing factor of the comprehensive causal weight of the previous time step; the total causal weight of the current node is calculated based on the comprehensive causal weights of the different edges connecting each node. If the total causal weight is less than the set threshold, the current node and the corresponding connected edge are removed. The total causal weight of each node is as follows: ； in represents the comprehensive causal weight of the edge between the kth node and the i-th node, represents the comprehensive causal weight of the edge between the kth node and the jth node, is the total causal weight of the current node.

2. The method for constructing a dynamic causal model for chronic disease management according to claim 1, characterized in that: A sliding window is set according to the time characteristics of the multimodal dynamic health data, and the multimodal dynamic health data is decomposed into long-term component features and short-term component features based on the sliding window.

3. The method for constructing a dynamic causal model for chronic disease management according to claim 1 is characterized in that: The conditional independence verification of multimodal dynamic health data is performed to remove irrelevant edges and obtain candidate edges. For the correlation between the two multimodal dynamic health data corresponding to all candidate edges, the candidate edges with correlation greater than the set threshold are retained as the edges of the edge set.

4. The method for constructing a dynamic causal model for chronic disease management according to claim 1, characterized in that: Calculate the marginal causal effect value of the current new multimodal dynamic health data and all the node sets in the dynamic causal model. When the marginal causal effect value is greater than a certain threshold, establish an edge between the new multimodal dynamic health data and the corresponding node, and use the correlation between the new multimodal dynamic health data and the associated nodes as the initial weight of the edge.

5. An application method of a dynamic causal model for chronic disease management, implemented according to any one of claims 1 to 4, characterized in that: The following steps are involved: Select the cause node and target node; Searching for potential edges from the cause node to the target node as potential dynamic causal chains, where the causal strength of the potential edges is greater than a set threshold; The potential dynamic causal chain with the greatest causal strength is selected as the dynamic causal chain.

6. An application method of a dynamic causal model for chronic disease management, implemented according to any one of claims 1 to 4, characterized in that: The following steps are involved: Select at least two intervention nodes and a target node; Searching for all influence edges from each intervention node to the target node, wherein the influence edges include direct edges from each intervention node to the target node, and indirect edges from the intervention node to the target node through other nodes; Calculate the influence of each intervention node based on all influencing edges and intervention changes of the intervention node; The intervention cost-effectiveness of each intervention node is calculated based on the influence of each intervention node and the intervention cost of the corresponding intervention node; Select intervention nodes with high cost-effectiveness.

7. The application method of the dynamic causal model for chronic disease management according to claim 6 is characterized in that: The direct influence of the current intervention node is calculated based on the direct edge, and the indirect influence of the current intervention node is calculated based on the indirect edge. The indirect influence and direct influence are combined as the influence of the current intervention node.

Images