Investigating explainable arrhythmia classification using class-specific ensemble model from segmented ECG signals.

Accurate arrhythmia classification from short clinical electrocardiograms (ECGs) remains challenging, particularly in terms of performance and interpretability. Prior studies are often limited to single-lead, use uniform model fusion, or report point estimates without uncertainty. To address this, we propose a novel class-specific weighted ensemble for 12-lead, 10 s ECG segments that fuses multiple models with per-class weights and provides lead-resolved SHapley Additive exPlanations (SHAP) explanations over 360 features. Four rhythms (sinus bradycardia (SB), sinus tachycardia (ST), atrial flutter (AF), supraventricular tachycardia (SVT)) were evaluated on the Chapman-Shaoxing dataset using an estimation-based framework with Wilson and Newcombe 95% confidence intervals (CIs) and Cohen'sh. Compared with a Bagging baseline, the proposed ensemble yields higher SB recall on both validation and test (Δ +0.040 and +0.050; CIs entirely positive; smallh), higher ST precision (Δ +0.070 and +0.080; smallh), and comparable overall accuracy (ΔCIs spanning 0;h≈ 0.00-0.02). AF and SVT differences are small or statistically imprecise. SHAP highlights lead-specific contributors (e.g. P-wave area in V2), indicating a path toward minimal-lead configurations while preserving performance. Collectively, integrating per-class weighted fusion, lead-level interpretability, and CI-based estimation on short 12-lead ECGs yields class-specific gains (higher SB sensitivity and ST precision), and SHAP pinpoints lead V2-with P-wave area among the top contributors. The outcome of this study provides potential insights that may be used to classify arrhythmia more accurately for each class and develop minimal lead configurations, given the growing use of wearable and portable devices. However, further studies and data are required for clinical application and device development.