Method for ischemic heart disease risk prediction based on continuum of care data

Abstract

The application discloses an ischemic heart disease risk prediction method based on continuous physical examination data and belongs to the field of biological information technology. In the application, continuous physical examination data and time interval data between each physical examination are vectorized, and time perception attention mechanism is used to perform representation learning on the two, so that different attention scores are given to each physical examination, and index data of the physical examination and the time interval data are fused. Subsequently, the most significant fusion features are introduced into a prediction layer for prediction, and a cross-entropy loss function is used to measure the relationship between the predicted value and the measured value. In order to improve the explainability of the classification of the method, Mendelian randomization analysis is used to analyze the causal relationship between part of the physical examination indexes and the ischemic heart disease, so as to enhance the reliability of the model in identifying disease risk factors. Compared with the traditional method, the application not only improves the performance of the ischemic heart disease risk prediction model, but also enhances the explainability of the model.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a method for predicting the risk of ischemic heart disease based on continuous physical examination data. Background Technology

[0002] Existing disease risk assessment methods typically employ techniques such as convolutional neural networks, recurrent neural networks, and machine learning to predict and identify disease risk factors.

[0003] However, existing methods for predicting disease risk require inputting single physical examination data into deep learning or machine learning models. These methods do not take into account the processing of continuous physical examination data, the impact of time interval differences between consecutive physical examinations on disease risk prediction, and the interpretability of the models. Summary of the Invention

[0004] To address the problems mentioned in the background section, this invention provides a method for predicting the risk of ischemic heart disease based on continuous physical examination data. This method solves the problems of existing technologies not considering the processing of continuous physical examination data, the impact of time interval differences between consecutive physical examinations on disease risk prediction, and the interpretability of the model.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for predicting the risk of ischemic heart disease based on continuous physical examination data includes the following steps:

[0007] Step 1: Collect medical record data and physical examination data including time data for the physical examination samples. Based on the medical record data, physical examination samples with ischemic heart disease are defined as positive samples, and the rest of the physical examination samples are defined as negative samples. The physical examination data is preprocessed and sorted by date to obtain the physical examination data sequence. The time interval matrix is ​​calculated through the time data. The dataset containing the training set, validation set and test set is constructed through the time interval matrix and the physical examination data sequence.

[0008] Step 2: Construct a model for prediction. Input the physical examination data sequence from the training set into a fully connected layer of the model to obtain the feature representation of the physical examination items. Combine the time interval matrix to generate the time interval feature representation. At the same time, construct the absolute position embedding representation. Based on the physical examination item feature representation, the time interval feature representation, and the absolute position embedding representation, calculate the fused feature sequence of the physical examination sample through a time-aware attention mechanism.

[0009] Step 3: Input the fused feature sequence into the prediction layer of the model, obtain the probability of ischemic heart disease through feature extraction and risk calculation, optimize the model parameters with the loss function, adjust the model using the validation set, evaluate the model performance and identify risk factors for ischemic heart disease through the test set.

[0010] Preferably, in step one, the preprocessing includes deleting physical examination samples with a missing rate of physical examination items greater than a preset proportion, fewer than a preset number of physical examinations, or outliers from the physical examination data; normalizing the retained physical examination data; when constructing the dataset, the total number of physical examination samples is N. First, the average number of physical examinations for all physical examination samples is calculated and rounded down to obtain a unified number of physical examinations n. For samples with a number of physical examinations less than n, zeros are inserted at the beginning of the sequence of physical examination data to fill in the gaps. For samples with a number of physical examinations greater than n, the last n physical examination data are taken to form the physical examination data sequence.

[0011] Preferably, in step two, the dimension of the fully connected layer is (h, d), where h is the number of physical examination items and d is the latent dimension of the model. The output of the fully connected layer is activated by an activation function to obtain the feature representation of the physical examination items. When generating the time interval feature representation, a maximum time interval k is set. If the time interval is greater than k, k is taken by default. The absolute position embedding representation is calculated and generated by creating an absolute position vector.

[0012] Preferably, in step two, the calculation process of the time-aware attention mechanism is as follows: the time-aware attention score is calculated using a compatibility function, the formula for which is:

[0013] ;

[0014] Where WQ and WK are both learnable parameter matrices of dimension d×d. M is the scaling factor, T is the embedded physical examination item feature representation, P is the time interval feature representation, and P is the absolute position embedding representation.

[0015] The attention score is converted into attention weights using the Softmax function, as shown in the formula:

[0016] ;

[0017] The fused feature sequence is obtained by calculating the weighted sum of the physical examination item feature representation, time interval feature representation, and absolute position embedding representation using attention weights.

[0018] Preferably, for the fused feature sequence, nonlinear features are assigned through a point-forward network, and after processing with layer normalization, residual connection and dropout techniques, the fused feature sequence is input into the prediction layer of the model.

[0019] Preferably, in step three, feature extraction adopts the max pooling method to extract the most significant feature value in the fused feature sequence. Risk calculation is implemented through a fully connected layer. First, the disease risk score is calculated, and then it is converted into the probability of ischemic heart disease through the Softmax function, and then output as the result of the model.

[0020] Preferably, in step three, the loss function is the cross-entropy loss function, with the following formula:

[0021] ;

[0022] Among them, yi, represents the fact of being diagnosed with ischemic heart disease and the predicted probability, respectively, where i is the number of physical examinations.

[0023] Preferably, in step three, when evaluating model performance, the calculated metrics include Precision value, Recall value, F1 score, Accuracy value, and AUC value. The specific calculation method is as follows:

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] Wherein, Precision represents the proportion of samples that the model predicts as positive but are actually positive; F1-score is the harmonic mean of Precision and Recall; TPR, or Recall, measures the model's ability to correctly identify positive samples; FPR represents the proportion of negative samples that are incorrectly identified as positive samples; Accuracy represents the proportion of samples correctly predicted by the model out of the total number of samples; TP represents the number of correctly predicted positive samples; FP represents the number of incorrectly predicted positive samples; TN represents the number of correctly predicted negative samples; FN represents the number of incorrectly predicted negative samples; and AUC, the area under the ROC curve, is used to evaluate the performance of the binary classifier model.

[0030] Preferably, in step three, when identifying risk factors for ischemic heart disease, a t-test is first performed on the physical examination data of positive and negative result samples in the test set. The p-value is used to determine whether there are significant differences between the two types of samples in specific physical examination indicators. Then, the attention coefficient matrix and physical examination data of the positive result samples are extracted and divided into the maximum attention coefficient group and other groups according to the attention score. A t-test is performed on each physical examination item of the two groups of physical examination data. Finally, attribution score analysis is performed on each physical examination item. The physical examination items with differences are combined with the items with positive attribution scores to determine the risk factors for ischemic heart disease.

[0031] Preferably, after identifying the risk factors for ischemic heart disease, the risk factors are used as exposure factors, and ischemic heart disease is used as an outcome variable for Mendelian randomization analysis to explain from a genetic perspective why the model focuses on specific physical examination items to predict the risk of disease.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] This application vectorizes continuous physical examination data and the time intervals between each examination, and uses a time-aware attention mechanism to learn representations of both, thereby assigning different attention scores to each examination. The examination indicator data and their time interval data are then fused, and the most significant fused features are imported into the prediction layer for prediction. A cross-entropy loss function is used to measure the relationship between predicted and measured values. To improve the interpretability of this classification method, this application also utilizes Mendelian randomization analysis to analyze the causal relationship between some physical examination indicators and ischemic heart disease, thereby enhancing the reliability of the model in identifying disease risk factors. Compared with traditional methods, this invention not only improves the performance of the ischemic heart disease risk prediction model but also enhances the model's interpretability. Attached Figure Description

[0034] Figure 1 This is a flowchart of the application;

[0035] Figure 2 This is a schematic diagram of a method for predicting the risk of ischemic heart disease based on continuous physical examination data.

[0036] Figure 3 This is a schematic diagram showing the comparison results between this application and existing methods;

[0037] Figure 4 This is a schematic diagram showing the t-test results of physical examination items for positive and negative samples.

[0038] Figure 5 This is a schematic diagram of the t-test results for each physical examination item for a positive result sample. Detailed Implementation

[0039] To facilitate understanding of the technical content of this invention by those skilled in the art, the invention will be further described in detail below with reference to the accompanying drawings and specific examples. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the scope of the invention.

[0040] Methods for predicting the risk of ischemic heart disease based on continuous physical examination data, such as Figure 1 , Figure 2 As shown, the specific steps include:

[0041] S1. Collect physical examination data and medical record data of physical examination samples, preprocess the two types of data and construct a dataset, wherein the dataset is divided into a training set, a validation set and a test set, specifically:

[0042] S11. Collect the physical examination data and medical record data of all physical examination samples, define samples with ischemic heart disease clearly confirmed in the medical record data as positive samples, and the rest as negative samples;

[0043] S12. Delete physical examination samples with a missing rate of physical examination item data greater than 95%, less than 2 physical examination times, or with abnormal values;

[0044] S13. Normalize all physical examination data, and sort the physical examination data of the physical examination samples by physical examination date;

[0045] S14. Extract the physical examination data of each physical examination sample , where is all the physical examination data of the physical examination sample at the i-th physical examination, m ij is the data of the j-th physical examination item of the sample at the i-th physical examination, t i is the time of the i-th physical examination, and N is the total number of physical examinations of the sample;

[0046] S15. Calculate the average number of physical examinations of all physical examination samples and round it down to be the unified number of physical examinations (n). For samples with the number of physical examinations less than n (N<n), we pad 0s in front of their physical examination data; for samples with the number of physical examinations greater than n (N>n), we take the last n pieces of physical examination data as the sample physical examination data:

[0047] ;

[0048] ;

[0049] S16. Perform calculation through the following formula to obtain a time interval matrix (t_matrix):

[0050] ;

[0051] S17. The time interval matrix t_matrix and the physical examination data sequence m_seq are used as the final sample data to construct a dataset, and the dataset is randomly divided into a training set, a validation set and a test set at a ratio of 7:1.5:1.5.

[0052] S2. Based on the obtained dataset, input the physical examination data and time data into the embedding layer of the model to realize data vectorization, specifically including:

[0053] ​S21. Input m_seq into the model, use a (h,d) fully connected layer to map m_seq from an n×h matrix to an n×d matrix, and use the ReLU function to give it non-linear characteristics, finally obtaining the model's physical examination item feature representation item_emb, where h is the number of physical examination items and d is the latent dimension of the model:

[0054] ;

[0055] S22. Create an absolute position vector and an absolute position embedding representation Where pos_seq represents the absolute position of each physical examination in the examination sequence, pos_emb is the embedding of pos_seq, R represents a real matrix, and the formula for calculating pos_emb is as follows: r is the dimension:

[0056] ;

[0057] S23. The time interval embedding matrix is... Where k represents the maximum time interval. If the time interval is greater than k, it defaults to k; otherwise, it remains unchanged. The embedded time interval feature is represented as follows:

[0058] ;

[0059] S3. Input the feature vector into the model's time-aware attention layer for representation learning, thereby fusing the physical examination data and time data to obtain the attention score for each physical examination, specifically including:

[0060] S31. The input physical examination data, time interval data, and location data are used to calculate a time-aware attention score using a compatibility function. The data will be labeled Q, K, and Y in the upper right corner to distinguish them as queries, keys, and values ​​used as attention mechanisms.

[0061] ;

[0062] in , Both are learnable parameter matrices of dimension d×d, and scale factors. To avoid excessively large inner product values, M is item_emb, T is time_emb, and P is pos_emb;

[0063] S32. Use the Softmax function to convert attention scores into attention weights:

[0064] ;

[0065] S33. Calculate the weighted sum of the physical examination data embedding, time interval embedding, and location embedding using attention weights, and use it as a new sequence representation of the physical examination sample. ,in The fusion features of each physical examination are represented as follows:

[0066] ;

[0067] S34. Imparting nonlinear characteristics to Z using a point-forward network:

[0068] ;

[0069] in , All four are learnable parameters;

[0070] S35. Use layer normalization, residual connection and dropout techniques to solve the problems of overfitting and training instability that may occur in the model, and output Z as a time-aware attention layer and input it into the prediction layer;

[0071] S4. Input the fused data into the prediction layer for risk prediction and train the model. Use the cross-entropy loss function to measure the relationship between the model's prediction layer output and the measured values ​​to determine the optimal parameters of the model. Specifically, this includes:

[0072] S41. Extract the most significant feature value from the fused features using max pooling;

[0073] S42. Utilize a fully connected layer to synthesize the last physical examination data z in Z. n Calculate the disease risk score:

[0074] ;

[0075] Among them W pred These are learnable parameters;

[0076] S43. Use the Softmax function to convert the risk score into the probability of the sample being diagnosed with ischemic heart disease at the next physical examination, and output it as the result of the model;

[0077] S43. Use the cross-entropy loss function as the objective function to optimize the parameters:

[0078] ;

[0079] Where y i , These represent the fact of being diagnosed with ischemic heart disease and the predicted probability, respectively.

[0080] The prediction result is defined as:

[0081] ;

[0082] in, , a and b represent the risk probabilities of the person undergoing the next physical examination not having ischemic heart disease and having ischemic heart disease, respectively, without human intervention. iIHD-TADP is the name of the entire model, which stands for Time-Aware Disease Prediction Model for Ischemic Heart Disease.

[0083] S5. Input the independent test set into the trained model to predict the risk of ischemic heart disease in the samples in the test set, specifically including:

[0084] S51. Train the model using the training and validation sets in the dataset, and save the trained model;

[0085] S52. Input the test set data into the trained model to make predictions and obtain the prediction results for the test set;

[0086] S53. Calculate the model's prediction accuracy using the prediction results and actual results from the test set.

[0087] Furthermore, the classification accuracy mentioned in step S53 includes Precision value, Recall value, F1 score, Accuracy value, and AUC value, which are calculated as follows:

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] In this context, Precision represents the proportion of actual positive samples that the model predicts as positive; F1-score is the harmonic mean of Precision and Recall; TPR, also known as Recall, measures the model's ability to correctly identify positive samples; FPR represents the proportion of negative samples that are incorrectly identified as positive samples; and Accuracy represents the proportion of correctly predicted samples out of the total number of samples. TP represents the number of correctly predicted positive samples, FP represents the number of incorrectly predicted positive samples, TN represents the number of correctly predicted negative samples, and FN represents the number of incorrectly predicted negative samples. AUC, the area under the ROC curve, is typically used to evaluate the performance of binary classifier models.

[0094] S6. By performing t-tests, attribution score analysis, and Mendelian randomization analysis on the data corresponding to the two types of prediction results in the independent test set, the explanation of the ischemic heart disease risk prediction method is completed, specifically including:

[0095] S61. Based on model prediction, the test set is divided into positive result samples and negative result samples, and the physical examination data corresponding to the two types of samples are obtained;

[0096] S62. Perform a t-test on the physical examination data of the two types of samples, and use the p-value to determine whether there are significant differences in specific physical examination indicators between the two types of samples.

[0097] S63. If there is a significant difference, it means that the model considers the physical examination item to be an important factor in distinguishing positive and negative samples, and it also theoretically proves that the model is possible.

[0098] S64. Extract the attention coefficient matrix and physical examination data of positive result samples, and divide the physical examination data into the maximum attention coefficient group and other groups according to the attention score of each physical examination;

[0099] S65. Perform a t-test on each physical examination item in the two groups of physical examination data, and use the p-value to obtain the physical examination items with significant differences. This result explains why the model assigns different weight coefficients to each physical examination.

[0100] S66. Perform attribution score analysis on each physical examination item. By removing the data of a certain physical examination item from the test set and inputting it into the same model, compare the difference in model performance before and after removing the physical examination item to obtain the importance of the physical examination item to the model, i.e., the attribution score of each physical examination item.

[0101] S67. By comparing the attribution scores of physical examination items, physical examination items that are conducive to the model's prediction of ischemic heart disease can be identified, and such physical examination items will also be regarded as risk factors for ischemic heart disease.

[0102] S68. Using the risk factors for ischemic heart disease identified in the above steps as exposure factors, and using ischemic heart disease as the outcome variable, perform Mendelian randomization analysis.

[0103] Among them, the differential physical examination items mentioned in S62 and S65 and the positive attribution scores in S66 can be regarded as risk factors for ischemic heart disease identified by the model. The results obtained in S66, combined with the results of S62 and S65, can obtain more accurate risk factors for ischemic heart disease. The Mendelian randomization results in S68 are used to explain from a genetic perspective why the model predicts the risk of ischemic heart disease by focusing on such physical examination items.

[0104] This application vectorizes continuous physical examination data and the time intervals between each examination, and uses a time-aware attention mechanism to learn representations of both, thereby assigning different attention scores to each examination and fusing the examination indicator data and their time interval data. The most significant fused features are then imported into the prediction layer for prediction, and the cross-entropy loss function is used to measure the relationship between predicted and measured values. To improve the interpretability of this method's classification, this invention combines the model's attention weights, t-tests, and attribution score analysis to analyze the model's classification basis and the disease risk factors it identifies. The t-test not only proves the theoretical feasibility of this invention but also, combined with the attention weights assigned to each examination, identifies important risk factors for ischemic heart disease. Attribution score analysis determines the importance of each examination indicator to this method. Furthermore, this method utilizes Mendelian randomization analysis to analyze the causal relationship between some examination indicators and ischemic heart disease, enhancing the reliability of the model's identification of disease risk factors. Compared with traditional methods, this invention not only improves the performance of the ischemic heart disease risk prediction model but also enhances its interpretability.

[0105] Figure 2The diagram illustrates the principle of a method for predicting ischemic heart disease risk based on continuous physical examination data and a time-aware attention mechanism, as provided in this embodiment of the invention. a) In the data processing stage, examination items with a missing rate >95% and individuals with fewer than two examinations are first removed. The data is then normalized and arranged chronologically to obtain a time series and a physical examination data matrix. The time interval matrix is ​​calculated using the time series. b) In the embedding layer, an absolute position embedding matrix is ​​created, and generalization and embedding operations are performed on the sample's physical examination data matrix and time interval matrix, respectively. c) A multi-head time-aware attention mechanism is used to synthesize the sample's physical examination data, absolute position of the examination, and time interval data to generate a comprehensive feature representation of each physical examination. The diagram shows the specific process of the time-aware attention mechanism, where a point-forward feedforward network imparts non-linear characteristics to the data. d) Max pooling is used to obtain the most significant comprehensive feature of the sample's last physical examination, which is then used for prediction.

[0106] Figure 3 The results show a comparison between the present invention and existing methods on the same dataset. To demonstrate the prediction accuracy of the iIHD-TADP method, we compared its performance with common CNN, RNN, and Transformer models on the same dataset. The iIHD-TADP model outperformed the other three models in terms of Recall, Accuracy, F1-score, and AUC, demonstrating the effectiveness of the present invention. Furthermore, the comparison between the iIHD-TADP method and the traditional Transformer method highlights the importance of time intervals in disease risk prediction.

[0107] Figure 4 The results shown are the t-test results of physical examination items for positive and negative samples. Figure 4As can be seen, in this embodiment, iIHD-TADP considers the following physical examination parameters as: Age, Fasting plasma glucose (FPG), Mean corpuscular volume (MCV), Systolic blood pressure (SBP), Diastolic blood pressure (DBP), Mean corpuscular hemoglobin (MCH), Red blood cell distribution width (RDW), Total cholesterol (TC), Body mass index (BMI), Creatinine (CRE), Low-density lipoprotein (LDL), Aspartate aminotransferase (AST), Absolute eosinophil count (EO#), and Eosinophil percentage (E The following parameters are used to distinguish between positive and negative samples: absolute monocyte count (MONO#), monocyte percentage (MONO%), absolute lymphocyte count (LYMPH#), neutrophil percentage (NEUT%), high-density lipoprotein (HDL), urine pH, mean corpuscular hemoglobin concentration (MCHC), total red blood cell count (RBC), absolute basophil count (BASO#), basophil percentage (BASO%), urine specific gravity (SG), and platelet count (PLT).

[0108] Figure 5 The results shown are the t-test results of the physical examination items for each positive result sample. Figure 5 It is evident that, among samples with positive results, the model pays more attention to changes in eight physical examination indicators: absolute basophil count (BASO#), basophil percentage (BASO%), monocyte percentage (MONO%), neutrophil percentage (NEUT%), absolute neutrophil count (NEUT#), lymphocyte percentage (LYMPH%), low-density lipoprotein (LDL), and total white blood cell count (WBC).

[0109] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the risk of ischemic heart disease based on continuous physical examination data, characterized in that, Includes the following steps: Step 1: Collect medical record data and physical examination data including time data for the physical examination samples. Based on the medical record data, physical examination samples with ischemic heart disease are defined as positive samples, and the rest of the physical examination samples are defined as negative samples. The physical examination data is preprocessed and sorted by date to obtain the physical examination data sequence. The time interval matrix is ​​calculated through the time data. The dataset containing the training set, validation set and test set is constructed through the time interval matrix and the physical examination data sequence. Step 2: Construct a model for prediction. Input the physical examination data sequence from the training set into a fully connected layer of the model to obtain the feature representation of the physical examination items. Combine the time interval matrix to generate the time interval feature representation. At the same time, construct the absolute position embedding representation. Based on the physical examination item feature representation, the time interval feature representation, and the absolute position embedding representation, calculate the fused feature sequence of the physical examination sample through a time-aware attention mechanism. In step two, the calculation process of the time-aware attention mechanism is as follows: the time-aware attention score is calculated using a compatibility function, the formula of which is: ； Among them W Q and W K Both are learnable parameter matrices of dimension d×d. M is the scaling factor, T is the embedded physical examination item feature representation, P is the time interval feature representation, and P is the absolute position embedding representation. The attention score is converted into attention weights using the Softmax function, as shown in the formula: ； The weighted sum of the physical examination item feature representation, time interval feature representation, and absolute position embedding representation is calculated using attention weights to obtain the fused feature sequence; n is the uniform number of physical examinations, which is calculated by rounding down the average number of physical examinations for all physical examination samples; k is the set maximum time interval. Step 3: Input the fused feature sequence into the prediction layer of the model, obtain the probability of ischemic heart disease through feature extraction and risk calculation, optimize the model parameters with the loss function, adjust the model using the validation set, evaluate the model performance and identify risk factors for ischemic heart disease through the test set; In step three, feature extraction uses max pooling to extract the most significant feature values ​​from the fused feature sequence. Risk calculation is achieved through a fully connected layer. First, a disease risk score is calculated, and then it is converted into the probability of ischemic heart disease through the Softmax function, which is then output as the result of the model.

2. The method for predicting the risk of ischemic heart disease based on continuous physical examination data according to claim 1, characterized in that, In step one, preprocessing includes deleting physical examination samples with missing data rates greater than a preset proportion, fewer than a preset number of physical examinations, or outliers; normalizing the retained physical examination data; when constructing the dataset, the total number of physical examination samples is N; for samples with fewer than n physical examinations, zeros are inserted at the beginning of the physical examination data sequence; for samples with more than n physical examinations, the last n physical examination data are taken to form the physical examination data sequence.

3. The method for predicting the risk of ischemic heart disease based on continuous physical examination data according to claim 2, characterized in that, In step two, the dimension of the fully connected layer is (h, d), where h is the number of physical examination items and d is the latent dimension of the model. The output of the fully connected layer is activated to obtain the feature representation of the physical examination items. When generating the time interval feature representation, if the time interval is greater than k, k is taken by default. The absolute position embedding representation is generated by creating an absolute position vector.

4. The method for predicting the risk of ischemic heart disease based on continuous physical examination data according to claim 3, characterized in that, For the fused feature sequence, nonlinear features are assigned through a point-forward network. After processing with layer normalization, residual connections and dropout techniques, the fused feature sequence is input into the prediction layer of the model.

5. The method for predicting the risk of ischemic heart disease based on continuous physical examination data according to claim 1, characterized in that, In step three, the loss function is the cross-entropy loss function, and the formula is: ； Among them, y i , represents the fact of being diagnosed with ischemic heart disease and the predicted probability, respectively, where i is the number of physical examinations.

6. The method for predicting the risk of ischemic heart disease based on continuous physical examination data according to claim 1, characterized in that, In step three, when evaluating model performance, the calculated metrics include Precision, Recall, F1 score, Accuracy, and AUC. The specific calculation methods are as follows: ； ； ； ； ； Wherein, Precision represents the proportion of samples that the model predicts as positive but are actually positive; F1-score is the harmonic mean of Precision and Recall; TPR, or Recall, measures the model's ability to correctly identify positive samples; FPR represents the proportion of negative samples that are incorrectly identified as positive samples; Accuracy represents the proportion of samples correctly predicted by the model out of the total number of samples; TP represents the number of correctly predicted positive samples; FP represents the number of incorrectly predicted positive samples; TN represents the number of correctly predicted negative samples; FN represents the number of incorrectly predicted negative samples; and AUC, the area under the ROC curve, is used to evaluate the performance of the binary classifier model.

7. The method for predicting the risk of ischemic heart disease based on continuous physical examination data according to claim 1, characterized in that, In step three, when identifying risk factors for ischemic heart disease, a t-test is first performed on the physical examination data of positive and negative samples in the test set. The p-value in the t-test is used to determine whether there are significant differences between the two types of samples in specific physical examination indicators. Then, the attention coefficient matrix and physical examination data of the positive samples are extracted and divided into the maximum attention coefficient group and other groups according to the attention score. A t-test is performed on each physical examination item of the two groups. Finally, attribution score analysis is performed on each physical examination item. The physical examination items with differences are combined with the items with positive attribution scores to determine the risk factors for ischemic heart disease.

8. The method for predicting the risk of ischemic heart disease based on continuous physical examination data according to claim 7, characterized in that, After identifying the risk factors for ischemic heart disease, we used these risk factors as exposure factors and ischemic heart disease as an outcome variable in a Mendelian randomization analysis to explain from a genetic perspective why the model focuses on specific health check-up items to predict the risk of disease.

Images