Hybrid Amplitude Ordinal Partition Networks for ECG Morphology Discrimination: An Application to PVC Recognition

Various algorithms have emerged for automatic electrocardiogram (ECG) analysis, focusing on discerning subtle changes in arrhythmic morphology, especially premature ventricular contractions (PVCs). While ordinal partition networks (OPNs) have been effective in analyzing real-world time series data, conventional OPNs primarily concentrate on local ECG shapes, overlooking amplitude-level information. This study introduces an innovative method, absolute amplitude OPNs (AAOPNs), which incorporates amplitude-level variations (coarse-grained ECG) into ordinal patterns. These AAOPNs encompass all conceivable ordinal patterns as nodes, connected based on temporal sequences. Ten network measures are extracted from both OPNs and AAOPNs, capturing the shape and level-based PVCs characteristics. These measures are integrated into hybrid amplitude OPNs (HAOPNs) to construct support vector machine (SVM)-based PVCs recognition models. Evaluated on three baseline databases [Massachusetts Institute of Technology-Beth Israel Hospital arrhythmia database (96587 beats), St. Petersburg Institute of Cardiological Technics database (156373 beats), and China Physiological Signal Challenge 2020 database (987209 beats)], the proposed models demonstrate F1-scores of 97.02%, 93.06%, and 91.03% across databases in class-oriented evaluation, and 94.57%, 87.96%, and 88.89% in subject-oriented evaluation. Notably, the model achieves PVCs scores of 53904 and 59201 on the China Physiological Signal Challenge 2020 test set for the two evaluations, affirming its efficacy in PVCs recognition. This framework provides a versatile approach for detecting morphology anomalies in various physiological signals, such as heart sounds and pulses.