Classification of Valvular Heart Diseases Using Nonlinear Heart Rate Variability Indices in Concurrent Electrocardiograms, Seismocardiograms, and Gyrocardiograms

Abstract

Valvular heart disease (VHD) remains a major clinical burden, while simple signal-based screening methods are limited. This study tested whether nonlinear heart rate variability (HRV) indices from concurrent electrocardiograms (ECG), seismocardiograms (SCG), and gyrocardiograms (GCG) can differentiate five VHD labels: aortic regurgitation (AR), aortic stenosis (AS), mitral regurgitation (MR), mitral stenosis (MS), and tricuspid regurgitation (TR). Because the public dataset contains only VHD patients, the task was lesion-oriented classification rather than healthy-control versus VHD screening. Nonlinear HRV features from interbeat intervals were classified with eight machine-learning models using grid search and 5-fold patient-level crossvalidation. Macro F1 score was prioritized over accuracy because several imbalanced tasks showed inflated accuracy. The best result was obtained for SCG-based AS classification with Linear SVM (accuracy: 0.780 ± 0.081, macro F1 score: 0.755 ± 0.109). Averaged across classifiers and VHDs, GCG achieved the highest mean macro F1 score (0.518), followed by SCG (0.503) and ECG (0.491). Nonlinear HRV indices therefore provide an interpretable benchmark for comparing ECG, SCG, and GCG in VHD-oriented classification.