Multi-agent-based interpretable cardiovascular health assessment method and system

Abstract

The invention discloses an interpretable cardiovascular health assessment method and system based on multiple agents, and belongs to the technical field of biomedical engineering, and the method comprises the steps: processing original data to obtain a first data set, dividing the first data set, correspondingly inputting different agents, and enabling the agents to output corresponding risk prediction and confidence results; aggregating the risk prediction output by the intelligent agent to obtain a comprehensive risk score, and aggregating the weight value of the feature vector and the corresponding dominance to calculate a feature importance score; sorting the feature importance scores to obtain a dominant factor list, generating intervention priority scores according to the feature importance scores and the mutation scores, matching the dominant factors with corresponding suggestion entries, and outputting suggestions. According to the method, multi-source data are integrated, multiple agents are designed for cooperative reasoning, trend analysis is carried out on time evolution of key risk factors on the basis, a potential risk mutation window period is identified, and time reference is provided for early intervention.

Description

Technical Field

[0001] The present invention belongs to the field of biomedical engineering technology, and in particular relates to an interpretable cardiovascular health assessment method and system based on multi-agents. Background Art

[0002] Atherosclerotic cardiovascular disease (ASCVD) is one of the leading causes of death and disability worldwide. Its early identification and risk assessment are extremely important for improving prognosis and guiding intervention strategies. Currently, in clinical practice, cardiovascular risk assessment is still mainly based on traditional statistical models, such as the China-PAR model and the Framingham risk score. These models are constructed based on epidemiological data and perform static predictions by inputting limited structured variables (such as age, gender, blood pressure, cholesterol levels, etc.). Although these methods have certain population applicability and statistical interpretability, they have shown obvious limitations in terms of personalized medicine, continuous health management, and the integration of new medical data.

[0003] First, such models often overlook the dynamic evolution of patients' health over the long term. Many key risk factors (such as changes in blood lipids, inflammation levels, and lifestyle fluctuations) change over time, yet traditional models assess them based solely on measurements at a single point in time. This results in a minimal ability to detect long-term trends and critical mutations, failing to meet clinical needs for proactive early warning. Second, existing models are severely limited in terms of data dimensionality. With the advancement of digital healthcare, wearable devices, mobile health applications, electronic medical record systems, and public health databases have enabled the provision of richer individual health data, including natural language descriptions of symptoms and medical history, continuous heart rate and sleep quality data, environmental and climatic factors, and genetic risk scores. However, most traditional models struggle to process this unstructured or high-dimensional, heterogeneous information, resulting in low data utilization and biased assessment conclusions. Third, in real-world clinical applications, doctors and patients alike often face the problem of uninterpretable model outputs: models simply output a risk score or level, lacking a transparent explanation of the underlying rationale behind the model's decisions. This hinders informed decision-making by doctors and limits patient engagement and compliance. Especially at the moment when deep learning models are gradually being introduced into medical risk control scenarios, the contradiction of "high performance but unexplainable" has become more prominent, becoming one of the biggest obstacles to the implementation of the model.

[0004] To this end, we propose a multi-agent based explainable cardiovascular health assessment method and system to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems of insufficient data utilization and inability to dynamically predict in the existing technology, and to propose an interpretable cardiovascular health assessment method and system based on multi-agent.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A multi-agent based interpretable cardiovascular health assessment method, including:

[0008] S1: input original data and process the original data to obtain the first data set;

[0009] The raw data includes structured medical data, unstructured text data, and wearable device data;

[0010] The structured medical data is derived from an existing electronic medical record system and includes various static physical signs and characteristics;

[0011] The unstructured text data comes from free-format medical records recorded by doctors, including behavior and medical history;

[0012] The wearable device data is collected by the patient wearing the device, including average heart rate, total number of steps and total sleep time;

[0013] S2: Divide the first data set into a first subspace, a second subspace, and a third subspace based on data sources;

[0014] The first subspace data is derived from structured medical data, and static vital sign features are input to the first agent;

[0015] The second subspace data is derived from unstructured text data, and behavior and medical history label features are input to the second agent;

[0016] The third subspace data is derived from wearable device data, and dynamic wearable data features are input to the third agent;

[0017] The intelligent agent outputs corresponding risk predictions and confidence results; the risk predictions output by all the intelligent agents are aggregated to obtain a comprehensive risk score;

[0018] S3: Calculating the structural dominance of the agent to the comprehensive risk score using an exponential distance function, where the structural dominance is used to measure the degree of closeness between the risk prediction output by the agent and the comprehensive risk score;

[0019] Obtaining weight values ​​of feature vectors input into the agent, aggregating the weight values ​​of the feature vectors with the corresponding dominance to obtain feature importance scores; sorting the feature importance scores to obtain a list of dominant factors;

[0020] S4: constructing a smoothed trend estimate for the dominant factor, wherein the smoothed trend estimate is calculated by an exponentially weighted moving average driven by the factor's explanatory weight; defining a mutation score function, calculating the mutation score of the dominant factor, and recording a mutation point when the mutation score is higher than a preset threshold;

[0021] S5: Map the risk level through the comprehensive risk score, generate an intervention priority score for the dominant factor of the mutation point based on the feature importance score and the mutation score, map the intervention level based on the intervention priority score, match the corresponding suggestion items to the dominant factor and output the suggestion.

[0022] Preferably, the static physical characteristics include age, gender, height, weight, systolic blood pressure, diastolic blood pressure, total cholesterol, HDL-C, and fasting blood glucose.

[0023] Preferably, the unstructured text data uses a clinical NLP model to extract entity information and output a binarized result.

[0024] Preferably, a cross-agent uncertainty penalty mechanism is introduced in the aggregation of the comprehensive risk score, and the output variance of the agents in the validation set is used as a measure of instability.

[0025] Preferably, the intelligent agent is constructed using a two-layer fully connected neural network and trained using historical annotated data.

[0026] Preferably, the mutation score function obtains the mutation score by aggregating the smoothed trend estimate trend residual, the dominant contribution and the comprehensive risk score fluctuation.

[0027] Preferably, the intervention level is determined based on the relationship between the intervention priority score and a preset threshold, and the rules are as follows:

[0028] When the intervention priority score is lower than the first threshold, no intervention is required;

[0029] When the intervention priority score is not lower than the first threshold and lower than the second threshold, a mild intervention is performed;

[0030] When the intervention priority score is not lower than the second threshold and lower than the third threshold, moderate intervention;

[0031] When the intervention priority score is higher than or equal to the third threshold, strong intervention is performed.

[0032] A multi-agent based explainable cardiovascular health assessment system, including:

[0033] a data preprocessing module, inputting raw data into the data preprocessing module and processing the raw data to obtain a first data set; the raw data includes structured medical data, unstructured text data, and wearable device data; the structured medical data is derived from an existing electronic medical record system and includes various static physical signs; the unstructured text data is derived from free-form medical records recorded by doctors, including behavior and medical history; the wearable device data is collected by a device worn by the patient and includes average heart rate, total number of steps, and total sleep duration;

[0034] a risk assessment module, wherein the risk assessment module divides the first data set into a first subspace, a second subspace, and a third subspace based on the data source; the first subspace data is derived from structured medical data, and static physical sign features are input to the first agent; the second subspace data is derived from unstructured text data, and behavior and medical history label features are input to the second agent; the third subspace data is derived from wearable device data, and dynamic wearable data features are input to the third agent; the agents output corresponding risk predictions and confidence results; and the risk predictions output by all the agents are aggregated to obtain a comprehensive risk score;

[0035] A dominant factor identification module is configured to calculate the structural dominance of the agent over the comprehensive risk score using an exponential distance function. The structural dominance is used to measure the degree of proximity between the risk prediction output by the agent and the comprehensive risk score. The module obtains the weight value of the feature vector input to the agent, aggregates the weight value of the feature vector with the corresponding dominance to obtain a feature importance score, and sorts the feature importance scores to obtain a list of dominant factors.

[0036] A trend identification module is configured to construct a smoothed trend estimate for the dominant factor, the smoothed trend estimate being calculated by an exponentially weighted moving average driven by the factor's explanatory weight; a mutation scoring function is defined to calculate the mutation score of the dominant factor, and when the mutation score is higher than a preset threshold, it is recorded as a mutation point;

[0037] A suggestion generation module maps the risk level through the comprehensive risk score, generates an intervention priority score for the dominant factor of the mutation point based on the feature importance score and the mutation score, maps the intervention level based on the intervention priority score, matches the dominant factor with the corresponding suggestion item, and outputs the suggestion.

[0038] In summary, the technical effects and advantages of the present invention are as follows: the present invention integrates multi-source health data including structured indicators, unstructured medical record texts, wearable device data and external environmental information, and through the design of multiple intelligent agents with clear role division, conducts parallel reasoning on ASCVD risks from different medical perspectives; and performs trend analysis on the time evolution of key risk factors, identifies potential risk mutation windows, and provides a time reference for early intervention. The evaluation results ultimately output by the system include not only the risk level, but also the impact path, contribution weight and change trend diagram of each factor, and proposes personalized health management recommendations in combination with historical data to support doctors in making evidence-based clinical decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flow chart of the method steps in the present invention;

[0040] Figure 2 Schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0042] like Figure 1 As shown in FIG, a multi-agent based interpretable cardiovascular health assessment method includes:

[0043] S1: input original data and process the original data to obtain the first data set;

[0044] The raw data includes structured medical data, unstructured text data, and wearable device data;

[0045] The structured medical data is derived from an existing electronic medical record system and includes various static physical signs and characteristics;

[0046] The unstructured text data comes from free-format medical records recorded by doctors, including behavior and medical history;

[0047] The wearable device data is collected by the patient wearing the device, including average heart rate, total number of steps and total sleep time;

[0048] S2: Divide the first data set into a first subspace, a second subspace, and a third subspace based on data sources;

[0049] The first subspace data is derived from structured medical data, and static vital sign features are input to the first agent;

[0050] The second subspace data is derived from unstructured text data, and behavior and medical history label features are input to the second agent;

[0051] The third subspace data is derived from wearable device data, and dynamic wearable data features are input to the third agent;

[0052] The intelligent agent outputs corresponding risk predictions and confidence results; the risk predictions output by all the intelligent agents are aggregated to obtain a comprehensive risk score;

[0053] S3: Calculating the structural dominance of the agent to the comprehensive risk score using an exponential distance function, where the structural dominance is used to measure the degree of closeness between the risk prediction output by the agent and the comprehensive risk score;

[0054] Obtaining weight values ​​of feature vectors input into the agent, aggregating the weight values ​​of the feature vectors with the corresponding dominance to obtain feature importance scores; sorting the feature importance scores to obtain a list of dominant factors;

[0055] S4: constructing a smoothed trend estimate for the dominant factor, wherein the smoothed trend estimate is calculated by an exponentially weighted moving average driven by the factor's explanatory weight; defining a mutation score function, calculating the mutation score of the dominant factor, and recording a mutation point when the mutation score is higher than a preset threshold;

[0056] S5: Map the risk level through the comprehensive risk score, generate an intervention priority score for the dominant factor of the mutation point based on the feature importance score and the mutation score, map the intervention level based on the intervention priority score, match the corresponding suggestion items to the dominant factor and output the suggestion.

[0057] The specific implementation steps include:

[0058] Step 1: Multi-source data input module

[0059] The goal of this step is to collect patient health-related data from three key data sources (structured medical data, unstructured text data, and wearable device data) and achieve standardized input through a unified preprocessing process, providing clean, complete, and uniformly structured data support for subsequent multi-agent collaborative modeling. This step not only requires the collection and integration of multi-source heterogeneous data, but also handles key operations such as missing values, text extraction, time series aggregation, and numerical standardization to ensure that the generated feature table has sufficient data quality, semantic integrity, and format compatibility.

[0060] Input data includes the following three categories:

[0061] Structured medical data comes from the hospital's electronic medical record (EMR) system and is obtained through the FHIR standard API interface. Fields include age, gender, height, weight, systolic blood pressure (SBP), diastolic blood pressure (DBP), total cholesterol, HDL-C, fasting blood glucose, etc.

[0062] Unstructured text data comes from free-format medical records recorded by doctors, including content such as chief complaint, current medical history, and past medical history, which are obtained through the text export interface provided by the EMR system;

[0063] Wearable device data is collected by devices worn by patients (such as Apple Watch, Fitbit, and Huawei bracelets), uploaded to the cloud via mobile apps, and retrieved daily through HealthKit or Google Fit API. Data fields include average heart rate, total steps, and total sleep duration.

[0064] All structured data is unified in field names and units according to field mapping rules. For example, all lipid values ​​are unified in mmol / L. After formatting, the data is merged into a DataFrame structure, with columns as variables and rows as patient samples. Missing values ​​in the structured data are imputed using the K-nearest neighbor interpolation algorithm. For example, if patient A's systolic blood pressure value is missing, the Euclidean distance between patient A and other patients based on age, BMI, blood glucose, and other characteristics is calculated. The K = 5 most similar patients are selected, and their mean systolic blood pressure value is used as the imputation value:

[0065]

[0066] in, represents the estimated systolic blood pressure of patient A, P k is the systolic blood pressure of the kth similar patient. The input features are normalized before distance calculation.

[0067] Unstructured text data uses a clinical NLP model (based on BERT fine-tuning) to extract entity information, such as smoking history, diabetes history, and family history of hypertension. The system uses the NER model to identify keywords and contextually determine semantics, outputting Boolean labels. For example, if "smoking a pack a day" appears in the text, the smoking label is set to 1; if "denies smoking," it is set to 0; if no information is found, the default value is 0.

[0068] Wearable device data is aggregated by patient ID and time, and the average value for the past seven days is extracted. Taking heart rate as an example, the device uploads the field "heart_rate_average" daily. We average the value for the past seven days to obtain:

[0069]

[0070] in, Indicates 7-day average heart rate, HRi is the average heart rate on day i, taken from the JSON field. The same method is used for step count and sleep duration.

[0071] All numerical characteristics (including SBP, DBP, blood glucose, TC, HDL-C, Step mean, sleep duration, etc.) perform z-score normalization to unify the input features into zero mean and unit variance. The normalization formula is:

[0072]

[0073] Where x is the original value, μ is the sample mean, σ is the sample standard deviation, and x′ is the normalized value used by the model. NLP labels remain binary and are not included in normalization.

[0074] After all processing is completed, the data is organized into a standardized DataFrame format, where each row represents a patient sample and the column fields are as follows:

[0075] Numeric fields: age, BMI, SBP, DBP, HDL-C, fasting blood sugar, Step count, sleep duration;

[0076] NLP tag fields: smoking history, presence of diabetes, and family history of hypertension.

[0077] The entire table has no missing values, no duplicate fields, consistent format, and field naming that aligns perfectly with the subsequent model. This table is passed to the multi-agent reasoning module in JSON format via the RESTful API interface.

[0078] Step 2: Multi-agent reasoning module

[0079] The goal of this step is to input the structured patient data (numerical and labeled) obtained in Step 1 into a collaborative risk assessment module composed of multiple agents, enabling continuous and personalized modeling and prediction of ASCVD risk levels. This module not only performs the core reasoning task but also serves as the structural support for the subsequent steps of extracting dominant factors and predicting temporal trends. Therefore, within the modeling architecture, we not only strive for predictive accuracy but also specifically design decoupled sub-models to maintain independent channels for each feature type during inference, facilitating subsequent interpretation and tracking.

[0080] To address the feature learning confusion problem caused by multi-source heterogeneous data, we divide the input features into three subspaces based on their sources, each corresponding to a dedicated agent:

[0081] Agent A: static physical characteristics → {age, BMI, SBP, DBP, HDL-C, fasting blood glucose};

[0082] Agent B: behavior and medical history labels → {smoking, diabetes, family hypertension};

[0083] Agent C:

[0084] Each agent is constructed using a two-layer fully connected neural network with the following structure:

[0085] Input layer → Dense(8)+ReLU→Dense(4)+ReLU→Dense(1)+Sigmoid, output the risk value r corresponding to its subspace i , where i∈{A,B,C}.

[0086] To prevent certain features (such as device data) from being overweighted in the training set due to poor data quality, we introduced a confidence adjustment regularization term into the final aggregation formula, which adaptively penalizes the output of each agent based on the prediction fluctuations of the validation set. The specific comprehensive risk prediction formula is as follows:

[0087]

[0088] in:

[0089] R is the final ASCVD risk score (range [0,1], which can be divided into low, moderate, and high levels);

[0090] r i is the risk prediction value of agent i;

[0091] α i is the weight of the agent, satisfying ∑α i =1, initially set to the mean, automatically optimized through training;

[0092] Var val (r i ) represents the output variance of the agent in the validation set (as a measure of instability);

[0093] λ is an adjustable parameter (such as 0.05) that controls the weight of the regularization term to prevent a single agent from dominating the global output.

[0094] Compared with the traditional weighted average, this formula innovatively adds a cross-agent uncertainty penalty mechanism, which is particularly suitable for scenarios in which the quality of data sources varies greatly in clinical practice. For example, wearable device data is highly volatile, which may cause jitter in the predicted value in the short term. Traditional averaging is likely to overestimate its value, and the introduction of Var val (r i ) item, the system will actively reduce the risk contribution of the model.

[0095] All agent models used the same training process, using supervised learning based on historical annotated data (with 10-year ASCVD risk labels), with the optimization objective being a weighted cross-entropy loss with a regularization term, using the Adam optimizer, a batch size of 128, and 50 epochs of training. Model parameters and α i The weights are jointly optimized during training so that the final R not only expresses the global risk but also maintains structural interpretability.

[0096] The output is a structured risk prediction result, including:

[0097] Comprehensive risk score R: represents the continuous score of the patient's ASCVD risk level;

[0098] The three sub-models output r A 、r B 、r C : For use in subsequent leading factor identification and dynamic trend analysis modules;

[0099] Confidence adjustment results of each agent Var val (r i ): used to explain model stability and risk credibility.

[0100] All results are packaged into a JSON structure, as shown below:

[0101]

[0102] "risk_score": corresponds to the final comprehensive score R;

[0103] "risk_components": output r of three groups of agents A 、r B 、r C , corresponding to static features, behavioral features, and wearable features respectively;

[0104] "risk_variance": Var*val(r A )、Var*val(r B ), Var val (r C ).

[0105] Step 3: Identification and interpretation of dominant factors

[0106] The goal of this step is to identify the key factors that dominate the ASCVD risk assessment in the current patient based on the collaborative risk inference results of the agent in the previous step, combined with the agent output structure and verification fluctuation information. Unlike the traditional model post-hoc interpretation, we directly use the structural decoupling result r in Step 2 A 、rB 、r C The synergistic difference between the total score R and the risk factor score R establishes a mechanism for "credible dominant factor attribution." This mechanism not only considers the static importance of factors in the model, but also integrates the dynamic deviation of risk contributions and output stability. This structural explanatory mechanism provides higher clinical credibility for AI-assisted diagnosis in medical scenarios.

[0107] Inputs include:

[0108] Comprehensive risk score R;

[0109] Each agent outputs r A 、r B 、r C ;

[0110] Output stability index Var of each model in the validation set val (r i );

[0111] The input feature vectors of each of the three sub-models (subsets of the output data in Step 1) x (A) 、x (B) 、x (C) .

[0112] First, we calculate the structural dominance D of each agent on the final risk score i , which measures how close its output is to the total score and adjusts it based on its confidence fluctuations. The details are as follows:

[0113]

[0114] D i represents the dominance of model i∈{A,B,C};

[0115] |Rr i | indicates the distance between its prediction and the total score;

[0116] Var val (r i ) is the output variance of the agent in the validation set;

[0117] γ controls the penalty rate of distance (e.g., γ=5), and β controls the strength of the stability penalty (e.g., β=10).

[0118] This formula combines an exponential distance function with a denominator stability penalty to give higher explanation trust weights to models whose outputs are not only close but also stable.

[0119] Next, for each agent’s internal feature x (i), we introduce a sparse attention mechanism within the structure and embed trainable weights in the last layer of the neural network This weight is optimized during training by minimizing the risk error and explanation sparsity, and the optimization goal is:

[0120]

[0121] BCE is the binary cross entropy loss, which is used to optimize the output of sub-model i;

[0122] The second item is Sparse regularization term, compressing weight vector The non-zero number of

[0123] λ is the sparse regularization coefficient (e.g., λ = 0.01), which is specifically used for explanation layer optimization.

[0124] The innovation lies in that this structure jointly optimizes the "explanatory feature weights" and the original modeling objectives, avoiding the problem of separation between the explanatory path and the modeling path (i.e., "explanatory drift") and ensuring that the feature scoring results are true and credible.

[0125] For the current sample, we obtain the weight value of each input feature from the three agents respectively Then multiply by their respective dominance D i Aggregate and get the final normalized feature importance score:

[0126]

[0127] S j is the original feature x j The dominant contribution to the current risk score R;

[0128] Z is the normalization factor, ensuring that all S j The sum is 1;

[0129] j traverses all original input features, a total of 12 (including label type and numerical type).

[0130] Get S j After that, we sort them in descending order by score, retain the top 5 as the dominant risk factors for the patient, and also output the dominance degree D of each agent. i In order to judge "which channel contributes most to the prediction" as a whole.

[0131] The output includes:

[0132] The list of dominant factors {x j ,S j}: For doctors to read and intervene;

[0133] Agent dominance vector {DA ,D B ,D C}: Assists in determining which dimension of data is mainly driving this prediction;

[0134] Example of JSON structure output:

[0135]

[0136]

[0137] "dominant_agents": Based on R, r i and Var val (r i ) calculated dominance D i ;

[0138] "top_factors": Final dominant factor explanation result S j .

[0139] Step 4: Modeling the trend of dominant factors and identifying mutation points

[0140] The goal of this step is to build on the current set of dominant factors output in Step 3 and, combined with historical data collected at multiple points in time, create a time series trend model for these dominant factor values. This step further identifies whether any "abnormal mutations" have occurred at any given point in time. This step enhances the system's ability to dynamically monitor individual cardiovascular risk, making the model interpretable not only within a single assessment but also capable of monitoring the trajectory of risk factor changes over time, enabling early detection and intervention.

[0141] For each dominant factor x j Construct its smoothed trend estimate The method uses an exponentially weighted moving average (EWMA) driven by the explanatory weights of the factors:

[0142]

[0143] in:

[0144] is based on the interpretability score S j The time decay weight of , α is the time decay constant (such as 0.5);

[0145] W is the window width (e.g., the most recent 3 to 5 records);

[0146] This structure combines factor importance with time decay to control trend modeling and focuses on intensity. The more important the factor, the smoother and more credible the trend line.

[0147] After obtaining the trend line, we identify whether there is a factor mutation at the current time point t and define the mutation score for:

[0148]

[0149] in:

[0150] The first item is the trend residual, the larger the error, the more significant the deviation from the trend.

[0151] The second item is the explanatory weight, which ensures that the key factors receive more attention;

[0152] The third item is the fluctuation of the global risk score, which is used to verify whether the mutation may affect the overall risk status.

[0153] when (preset threshold, such as 0.2), the system regards it as a "mutation point" and records it for reference in subsequent intervention modules.

[0154] The following are examples:

[0155] If a patient wears a bracelet and uploads sleep duration for nearly 4 consecutive months, the results are 7.2h, 7.1h, 6.9h, and 5.3h respectively, and the explanatory score of this factor is S j =0.18, at this time the overall risk score of the system is R (t) If it increases by 0.15 compared to the previous time, then we get:

[0156]

[0157] If the threshold θ is set to 0.15, the system will judge that "the sleep duration has changed suddenly and is synchronized with the increase in overall risk", which is a suspicious mutation.

[0158] The output is structured as follows:

[0159] Trend estimates for each dominant factor and mutation score Warning information;

[0160] Output format examples are:

[0161]

[0162] "current_value":

[0163] "change_score": mutation score

[0164] "alert": warning information obtained based on threshold judgment

[0165] Step 5: Result output and intervention recommendations

[0166] The goal of this step is to integrate the multi-agent risk scoring results, ranking of dominant factors, and mutation point identification results constructed in steps two through four into structured risk level prompts and personalized intervention recommendations for doctors and patients. This module completes the system's "prediction-explanation-intervention" closed loop, requiring all outputs to be uniformly structured, variable-closed, and reproducible.

[0167] First, the system maps the current risk level according to the risk score R:

[0168]

[0169] At the same time, the credibility of the risk results is judged by using the maximum model fluctuation value max(Var val (r i If the value is < 0.05, it is labeled as "stable prediction"; if it is ≥ 0.1, it is labeled as "large fluctuation"; otherwise, it is labeled as "credible prediction".

[0170] For each dominant factor marked as mutated (alert j =true), the system scores S according to its explanatory power. j Mutation scoring Generate intervention priority scores:

[0171]

[0172] in:

[0173] S j Score the importance of the factor in the model structure;

[0174] Score its mutation at the current time point to measure the suddenness of its change;

[0175] I j It represents the final intervention intensity score, ranging from [0,1]. The higher the score, the more intervention is needed.

[0176] System according to I j Value Mapping Intervention Levels:

[0177] I j <0.1: no intervention required;

[0178] 0.1≤I j <0.3: mild intervention;

[0179] 0.3≤I j <0.6: moderate intervention;

[0180] I j ≥0.6: strong intervention.

[0181] The system then matches intervention content from a pre-set factor suggestion library. Each factor corresponds to a different level of suggestion items. For example:

[0182] Moderate intervention for "sleep duration": It is recommended to sleep no less than 7 hours per night and avoid staying up late continuously;

[0183] For strong intervention of "SBP": it is recommended to reduce sodium intake and monitor blood pressure for 7 consecutive days.

[0184] The technical solutions in the above-mentioned embodiments of the present application have at least the following technical effects or advantages: the present invention integrates multi-source health data including structured indicators, unstructured medical record texts, wearable device data and external environmental information, and through the design of multiple intelligent agents with clear role division, conducts parallel reasoning on ASCVD risk from different medical perspectives; and performs trend analysis on the time evolution of key risk factors, identifies potential risk mutation windows, and provides a time reference for early intervention. The evaluation results ultimately output by the system include not only the risk level, but also the impact path, contribution weight and change trend diagram of each factor, and combines historical data to propose personalized health management recommendations to support doctors in making evidence-based clinical decisions.

[0185] The present application also provides an interpretable cardiovascular health assessment system based on multi-agents, such as Figure 2 Shown, including:

[0186] a data preprocessing module, inputting raw data into the data preprocessing module and processing the raw data to obtain a first data set; the raw data includes structured medical data, unstructured text data, and wearable device data; the structured medical data is derived from an existing electronic medical record system and includes various static physical signs; the unstructured text data is derived from free-form medical records recorded by doctors, including behavior and medical history; the wearable device data is collected by a device worn by the patient and includes average heart rate, total number of steps, and total sleep duration;

[0187] a risk assessment module, wherein the risk assessment module divides the first data set into a first subspace, a second subspace, and a third subspace based on the data source; the first subspace data is derived from structured medical data, and static physical sign features are input to the first agent; the second subspace data is derived from unstructured text data, and behavior and medical history label features are input to the second agent; the third subspace data is derived from wearable device data, and dynamic wearable data features are input to the third agent; the agents output corresponding risk predictions and confidence results; and the risk predictions output by all the agents are aggregated to obtain a comprehensive risk score;

[0188] A dominant factor identification module is configured to calculate the structural dominance of the agent over the comprehensive risk score using an exponential distance function. The structural dominance is used to measure the degree of proximity between the risk prediction output by the agent and the comprehensive risk score. The module obtains the weight value of the feature vector input to the agent, aggregates the weight value of the feature vector with the corresponding dominance to obtain a feature importance score, and sorts the feature importance scores to obtain a list of dominant factors.

[0189] A trend identification module is configured to construct a smoothed trend estimate for the dominant factor, the smoothed trend estimate being calculated by an exponentially weighted moving average driven by the factor's explanatory weight; a mutation scoring function is defined to calculate the mutation score of the dominant factor, and when the mutation score is higher than a preset threshold, it is recorded as a mutation point;

[0190] A suggestion generation module maps the risk level through the comprehensive risk score, generates an intervention priority score for the dominant factor of the mutation point based on the feature importance score and the mutation score, maps the intervention level based on the intervention priority score, matches the dominant factor with the corresponding suggestion item, and outputs the suggestion.

[0191] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A multi-agent based interpretable cardiovascular health assessment method, characterized by: include: S1: input original data and process the original data to obtain the first data set; The raw data includes structured medical data, unstructured text data, and wearable device data; The structured medical data is derived from an existing electronic medical record system and includes various static physical signs and characteristics; The unstructured text data comes from free-format medical records recorded by doctors, including behavior and medical history; The wearable device data is collected by the patient wearing the device, including average heart rate, total number of steps and total sleep time; S2: Divide the first data set into a first subspace, a second subspace, and a third subspace based on data sources; The first subspace data is derived from structured medical data, and static vital sign features are input to the first agent; The second subspace data is derived from unstructured text data, and behavior and medical history label features are input to the second agent; The third subspace data is derived from wearable device data, and dynamic wearable data features are input to the third agent; The intelligent agent outputs corresponding risk predictions and confidence results; the risk predictions output by all the intelligent agents are aggregated to obtain a comprehensive risk score; S3: Calculating the structural dominance of the agent to the comprehensive risk score using an exponential distance function, where the structural dominance is used to measure the degree of closeness between the risk prediction output by the agent and the comprehensive risk score; Obtaining a weight value of a feature vector input into the agent, and aggregating the weight value of the feature vector with the corresponding dominance to obtain a feature importance score; Sorting the feature importance scores to obtain a list of dominant factors; S4: constructing a smoothed trend estimate for the dominant factor, wherein the smoothed trend estimate is calculated by an exponentially weighted moving average driven by the factor's explanatory weight; defining a mutation score function, calculating the mutation score of the dominant factor, and recording a mutation point when the mutation score is higher than a preset threshold; S5: Map the risk level through the comprehensive risk score, generate an intervention priority score for the dominant factor of the mutation point based on the feature importance score and the mutation score, map the intervention level based on the intervention priority score, match the corresponding suggestion items to the dominant factor and output the suggestion.

2. The multi-agent based interpretable cardiovascular health assessment method according to claim 1, characterized in that: The static physical characteristics include age, gender, height, weight, systolic blood pressure, diastolic blood pressure, total cholesterol, HDL-C, and fasting blood glucose.

3. The multi-agent based interpretable cardiovascular health assessment method according to claim 1, characterized in that: The unstructured text data uses a clinical NLP model to extract entity information and output a binarized result.

4. The multi-agent based interpretable cardiovascular health assessment method according to claim 1, characterized in that: A cross-agent uncertainty penalty mechanism is introduced in the aggregation of the comprehensive risk score, and the output variance of the agent in the validation set is used as a measure of instability.

5. The multi-agent based interpretable cardiovascular health assessment method according to claim 1, characterized in that: The intelligent agent is constructed using a two-layer fully connected neural network and is trained using historical annotated data.

6. The multi-agent based interpretable cardiovascular health assessment method according to claim 1, characterized in that: The mutation score function obtains the mutation score by aggregating the smoothed trend estimate trend residual, dominant contribution and comprehensive risk score fluctuation.

7. The multi-agent based interpretable cardiovascular health assessment method according to claim 1, characterized in that: The intervention level is determined based on the relationship between the intervention priority score and the preset threshold, and the rules are as follows: When the intervention priority score is lower than the first threshold, no intervention is required; When the intervention priority score is not lower than the first threshold and lower than the second threshold, a mild intervention is performed; When the intervention priority score is not lower than the second threshold and lower than the third threshold, moderate intervention; When the intervention priority score is higher than or equal to the third threshold, strong intervention is performed.

8. A multi-agent based explainable cardiovascular health assessment system, characterized by: include: a data preprocessing module, inputting raw data into the data preprocessing module and processing the raw data to obtain a first data set; the raw data includes structured medical data, unstructured text data, and wearable device data; the structured medical data is derived from an existing electronic medical record system and includes various static physical signs; the unstructured text data is derived from free-form medical records recorded by doctors, including behavior and medical history; the wearable device data is collected by a device worn by the patient and includes average heart rate, total number of steps, and total sleep duration; a risk assessment module, wherein the risk assessment module divides the first data set into a first subspace, a second subspace, and a third subspace based on the data source; the first subspace data is derived from structured medical data, and static physical sign features are input to the first agent; the second subspace data is derived from unstructured text data, and behavior and medical history label features are input to the second agent; the third subspace data is derived from wearable device data, and dynamic wearable data features are input to the third agent; the agents output corresponding risk predictions and confidence results; and the risk predictions output by all the agents are aggregated to obtain a comprehensive risk score; a dominant factor identification module, wherein the dominant factor identification module calculates the structural dominance of the agent to the comprehensive risk score using an exponential distance function, wherein the structural dominance is used to measure the degree of proximity between the corresponding risk prediction output by the agent and the comprehensive risk score; Obtaining a weight value of a feature vector input into the agent, and aggregating the weight value of the feature vector with the corresponding dominance to obtain a feature importance score; Sorting the feature importance scores to obtain a list of dominant factors; A trend identification module is configured to construct a smoothed trend estimate for the dominant factor, the smoothed trend estimate being calculated by an exponentially weighted moving average driven by the factor's explanatory weight; a mutation scoring function is defined to calculate the mutation score of the dominant factor, and when the mutation score is higher than a preset threshold, it is recorded as a mutation point; A suggestion generation module maps the risk level through the comprehensive risk score, generates an intervention priority score for the dominant factor of the mutation point based on the feature importance score and the mutation score, maps the intervention level based on the intervention priority score, matches the dominant factor with the corresponding suggestion item, and outputs the suggestion.

Images