Chest pain risk layering method and system with interpretability

Abstract

The invention provides an interpretable chest pain risk layering method and system. The method comprises the following steps: acquiring an electrocardiogram data set; performing data preprocessing on the electrocardiogram data set; based on the data set obtained after processing, utilizing a deterministic learning algorithm to extract the dynamic characteristics of the electric signals in the data set center, and constructing an electrocardio dynamic graph data set; constructing a chest pain risk layered characteristic variable data set by combining the static characteristic variable of the basic clinical information and the dynamic characteristic variable in the electrocardio dynamic graph data set; according to the chest pain risk hierarchical feature variable data set, calculating the marginal contribution of each feature variable to high-risk chest pain recognition through a random forest and an SHAP value; and in combination with the marginal contribution and expert experience, scores of all the feature variables are determined, all the feature variables are fused by using the scores, a total score is obtained through calculation, and corresponding chest pain risk layering grades are given according to the total score. According to the method, accurate risk stratification of the chest pain patient can be quickly realized in a short time, the misdiagnosis rate is reduced, and excessive diagnosis and treatment and medical resource waste are reduced.

Claims

1. An interpretable method for chest pain risk stratification, characterized by: The following steps are involved: Get the ECG dataset; performing data preprocessing on the electrocardiogram dataset; Based on the processed data set, a deterministic learning algorithm is used to extract the dynamic characteristics of the electrocardiogram signal in the data set and construct an electrocardiogram data set; Combining the static characteristic variables of basic clinical information with the dynamic characteristic variables in the electrocardiogram dataset, a chest pain risk stratification characteristic variable dataset was constructed. Based on the chest pain risk stratification characteristic variable dataset, the marginal contribution of each characteristic variable to identifying high-risk chest pain was calculated using random forest and SHAP values; The marginal contribution and expert experience are combined to determine the score of each characteristic variable, and the score is used to fuse the characteristic variables to calculate the total score, and the corresponding chest pain risk stratification level is given according to the total score.

2. The interpretable chest pain risk stratification method according to claim 1, wherein: The electrocardiogram dataset is a twelve-lead standard electrocardiogram with a set sampling rate; Alternatively, the basic clinical information includes the patient's gender, age, medical history of stroke, history of myocardial infarction / revascularization, history of hypertension, and history of diabetes.

3. The interpretable chest pain risk stratification method according to claim 1, wherein: The data preprocessing includes outlier removal and filtering noise reduction, wherein: The outlier elimination process includes removing the noise exceeding the set value and the incomplete 12-lead ECG data; The filtering and denoising process involves using smoothing to denoise the electrocardiogram, and reducing high-frequency noise in the signal by performing weighted averaging on the data near each data point.

4. The interpretable chest pain risk stratification method according to claim 1, wherein: The process of constructing the electrocardiogram dataset includes: The ECG signal is regarded as a non-stationary signal with periodic or recurrent characteristics generated by the complex nonlinear dynamic system of the heart, and is converted into a three-lead ECG vector signal; The vector cardiogram (ECG) signal is sampled and the sampled data corresponding to the ST-T segment of the ECG signal is intercepted. A deterministic learning algorithm based on the sampled data is used to perform local accurate neural network modeling of the internal system dynamics of the ST-T loop, thereby obtaining the internal dynamic characteristics of the ST-T loop of the ECG vector cardiogram (ECG) signal. Perform three-dimensional visualization of the above dynamic modeling results to obtain an electrocardiogram; For each electrocardiogram, the spatial discreteness of the electrocardiogram was extracted using the Lyapunov index, and the temporal discreteness of the electrocardiogram was obtained using Fourier transform, which was used as the electrocardiogram characteristic vector to describe the patient's myocardial ischemia state.

5. The interpretable chest pain risk stratification method according to claim 4, characterized in that: The ECG signal is regarded as a non-stationary signal with periodic or recurrent characteristics generated by the complex nonlinear dynamic system of the heart. The process of converting it into a three-lead ECG vector signal includes: the non-stationary signal is: Where X(t)=[x2(t),x2(t),...,x 22 (t)] T is the system state, representing the standard twelve-lead telecardiogram of the human body, F(X(t))=[f1(x(t)),f2(x(t)),...,f 12 (x(t))] T It is an unknown nonlinear dynamic; Convert to three-lead ECG vector signal: Where V(t) = [v1(t), v2(t), v3(t)] T ∈R 3 Represents the three-dimensional electrocardiographic vector signal, F v (V(t))=[F v1 (V(t)),F v2 (V(t)),F v3 (V(t))] is the nonlinear system dynamics, representing the inherent dynamic characteristics of the ECG vector signal.

6. The interpretable chest pain risk stratification method according to claim 4, wherein: The process of locally accurate neural network modeling of the ST-T loop internal system dynamics using a deterministic learning algorithm based on sampled data includes dynamic characteristics for: The above kinetic modeling results Perform three-dimensional visualization to obtain the electrocardiogram.

7. The interpretable chest pain risk stratification method according to claim 1, wherein: The process of calculating the marginal contribution of each feature variable to identifying high-risk chest pain using random forest and SHAP values ​​includes: Using chest pain risk level as the target variable, basic clinical information and electrocardiogram feature data were input to construct a random forest model; When splitting each node, m feature variables are randomly selected as candidate feature variables, feature variable selection is performed, these m feature variables are evaluated, and an optimal feature variable is selected for node division; Construct N decision trees to form a random forest; The prediction results of the model are explained by combining the random forest model with the SHAP value.

8. The interpretable chest pain risk stratification method according to claim 7, wherein: When evaluating characteristic variables, the Gini index is used as the evaluation criterion; Alternatively, the process of explaining the prediction results of the model by combining the random forest model with the SHAP value includes: For the random forest model, the SHAP value calculates the marginal contribution of each feature variable to the model output. The specific formula is: Where: φ j represents the SHAP value of feature variable j, F represents the set of all feature variables, S represents the subset of feature variables that does not include j, and f(S) represents the predicted value of the model when only feature variable S is included.

9. The interpretable chest pain risk stratification method according to claim 7, wherein: When calculating the scores of each characteristic variable, the specific process includes taking the characteristic variable with the smallest marginal contribution as the benchmark, calculating the relative scores of each characteristic variable, and then grading and assigning values ​​to age, electrocardiogram, and hs-cTNI based on clinical experience.

10. An interpretable chest pain risk stratification system, characterized by: include: a data acquisition module configured to acquire an electrocardiogram dataset; a data preprocessing module, configured to perform data preprocessing on the electrocardiogram dataset; The feature extraction module is configured to extract the dynamic features of the electrocardiogram signal in the data set based on the processed data set by using a deterministic learning algorithm to construct an electrocardiogram data set; A feature set construction module is configured to combine static feature variables of basic clinical information with dynamic feature variables in the electrocardiogram dataset to construct a chest pain risk stratification feature variable dataset; A marginal contribution calculation module is configured to calculate the marginal contribution of each characteristic variable to identifying high-risk chest pain through random forest and SHAP value based on the chest pain risk stratification characteristic variable data set; The chest pain risk stratification module is configured to combine the marginal contribution and expert experience to determine the score of each characteristic variable, use the score to fuse the various characteristic variables, calculate the total score, and give the corresponding chest pain risk stratification level according to the total score.

Images