Machine learning-based CAD detection using integrated ECG and PCG parameter features.

The combined analysis of electrocardiogram (ECG) and phonocardiogram signals(PCG) has demonstrated significant potential in the non-invasive detection of coronary artery disease (CAD). The efficacy of combining cardiac pathological parameters such as electromechanical delay (EMD), left ventricular ejection time (LVET), pre-ejection period (PEP), and the cardiac dysfunction index (SDI score) for the intelligent diagnosis of CAD remains unverified. In this study, an improved Pan-Tompkins algorithm was employed to locate the QRS complexes in the ECG signal accurately. A frequency-domain windowing threshold segmentation method was proposed to detect the S1, S2, and S3 peaks in the PCG signal. The automatic parameter extraction algorithm was then applied to compute four-time series-RR interval, EMD, LVET, and PEP along with the SDI score. In the feature extraction phase, a combination of time-domain, frequency-domain, and nonlinear measurement methods was employed. Comparative experiments were conducted using two support vector machine classification models (SVM-1, SVM-2) and two XGBoost classification models (XGBoost-1, XGBoost-2), each trained with different input features. Experimental results showed that including EMD, LVET, and PEP time-domain features significantly enhanced the classification performance of both the SVM and XGBoost models. The accuracy, sensitivity, specificity, and AUC metrics were superior to models using traditional features, with accuracy improving by 21% and 23%, respectively. The results of feature importance analysis combined with paired-sample Wilcoxon signed-rank tests further demonstrated the critical role of integrating EMD, PEP, and LVET in non-invasive CAD detection, with PEP showing the highest feature weight. These findings strongly validate the effectiveness of combining ECG and PCG parameters and features for the intelligent diagnosis of CAD.