Feature Engineering Design And Interpretation Of ECG Signals

Objective: This study aims to find the effective features associated with diseases through ECG characteristic engineering technology,correlate the features used in the diagnosis of various diseases with clinical results,obtain the medical relationship between the features and disease diagnosis,complete the interpretability analysis of artificial intelligence model,enhance the reliability of the clinical application of artificial intelligence model,and make contributions to the promotion of automatic ECG diagnosis to the clinic.Methods: Static electrocardiogram data from the cardiovascular department of a hospital were selected,including 28 types of abnormal cardiovascular diseases.A total of 893 features including statistical characteristics,nonlinear characteristics,time-domain characteristics,signal factor characteristics,traditional clinical characteristics of ECG,and potential co-occurrence matrix characteristics were constructed.By Lasso regression,shap value and Pearson correlation analysis were used for feature screening to obtain effective features of each disease.In this study,logistic regression,support vector machine,catboost and xgboost were used as classifiers to compare the diagnostic performance among effective features based on screening,traditional ECG clinical features and all feature sets,and then compared with the deep learning model based on original ECG data to verify the effectiveness of screening features.The correlation and contribution between features and diagnostic results were obtained by using shap interpretability method,and then the medical relationship between features and diseases was explained from the clinical perspective,so as to construct the framework of ECG characteristics.Finally,based on the ECG characteristic framework,6 types of abnormal cardiovascular diseases were used,and other data sets were used to verify the practicability of the ECG characteristic framework.Results: In this study,the effective characteristics of 28 types of cardiovascular diseases were screened,and the accuracy and recall rate indexes of the effective characteristics screened for each disease were above 95% in the optimal model.Moreover,these diseases require very few features.Except for several diseases such as atrial premature beat and ventricular premature beat,the features used in most diseases are reduced from nearly 900 to less than 20,which basically can greatly reduce the calculation cost.At the same time,in the diagnosis of some diseases such as ventricular premature beat,atrial premature beat,borderline premature beat,and ventricular presystole,the classification accuracy of the classification model based on traditional ECG characteristics is basically lower than 80%,and the accuracy of the classification model is directly increased to more than 95% by using the effective features of screening.The overall accuracy of classification performance of 6 classification models based on cardiac feature framework is up to 96.09%.Conclusion: In this study,multiple features such as time-frequency domain features,nonlinear features,statistical features and traditional ECG clinical features were extracted.Based on the theory of feature engineering and the interpretability analysis of shap,the validity of screening features was verified,and the interpretability analysis of these effective features was carried out from the aspect of clinical significance.The results of this study can promote the clinical application of automatic ECG diagnosis,improve the reliability of doctors and researchers on automatic ECG diagnosis,and provide reference value for subsequent research.