Detection of coronary microvascular dysfunction based on machine learning algorithm with multidimensional temporal-spatial features from electrocardiogram

Coronary microvascular dysfunction (CMD) causes myocardial ischemia and is associated with adverse cardiovascular events. This study explored the value of multidimensional electrocardiogram (ECG) features in capturing pathological changes linked to CMD. Global electrical heterogeneity (GEH) parameters were extracted from vectorcardiogram (VCG) signals synthesized from ECGs o 82 CMD patients and 252 controls. Temporal- and spatial-heterogeneity indices were extracted from VCG signals and cardiodynamicsgrams (CDGs), which were derived from VCGs' ST-T loops via deterministic learning algorithms. Entropy metrics were calculated across VCG waveforms, ECG signals, their ST-T segments, and CDGs. Feature selection was performed using sequential backward selection (SBS) and random forest (RF). The optimized feature sets were separately input into multilayer perceptron, support vector machine (SVM), XGBoost, and K-nearest neighbor algorithms to select the optimal model for CMD identification. Finally, the SVM models outperformed other models using both SBS- and RF-selected feature sets. The SVM model with SBS-selected features, comprising 5 GEH parameters, 3 VCG-based features, 1 CDG-based metrics, and 9 ECG-derived features, demonstrated superior discriminative performance compared to that with RF-selected features. It achieved high accuracy (0.923), specificity (0.925), sensitivity (0.917), and an area under the curve (0.970) in CMD classification. Generalization validation were performed on the 122 healthy ECGs from the Physikalisch-Technische Bundesanstalt diagnostic ECG database and 134 healthy ECGs from China Physiological Signal Challenge 2018 database further supported the superiority SVM model. It achieved high specificity values (>0.85) across both datasets. In conclusions, the integration of multidimensional ECG, VCG, CDG, and GEH features enhances the diagnostic performance of SVM model in detecting CMD. The developed methodology provides a non-invasive diagnostic tool with high sensitivity and specificity, showing promising prospects for potential clinical application.