Hybrid PCA-PINN framework for accurate and transparent ECG arrhythmia detection

Automated arrhythmia screening from electrocardiograms (ECGs) has advanced rapidly, yet many accurate systems lack physiological interpretability. We introduce a hybrid, physics augmented pipeline that fuses 50 statistical components from principal component analysis (PCA) with two interpretable, physics informed neural network (PINN) coefficients (c1, c3) estimated per image by minimizing a simplified CLG style partial differential equation residual. The fused 52 dimensional vector is classified by a lightweight multilayer perceptron (128 → 64 → 32, ReLU, softmax). On a patient disjoint held out test set of 1,161 ECG images (class counts: F=161; M/N/Q/S/V=200 each), the hybrid model attains 95.43% accuracy and 95.00% macro F1, outperforming a PCA only baseline (92.59% accuracy; 93.00% macro F1) and a PINN only variant (18.52% accuracy). Per class gains are most pronounced for clinically challenging supraventricular (S) and ventricular (V) ectopic beats (F1: 88% → 92% and 90% → 93%, respectively). One vs rest ROC analysis shows near perfect discrimination (AUC ≈ 1.00 for F/M/N/Q and ≈ 0.99 for S/V), while confusion matrix inspection confirms reduced misclassification within the S↔V pair (net 19 → 18). Qualitative exemplars indicate that c1 (propagation/relaxation) and c3 (damping/cross coupling) inject mechanistic context that complements shape dominant PCA features, aiding separation when morphology alone is ambiguous (e.g., widened QRS supraventricular beats or narrow complex ventricular ectopy). The resulting model is compact, interpretable, and competitive with deeper vision based pipelines, suggesting a practical path toward clinically transparent ECG classification and a template for physics aware learning in biomedical signal analysis.