Multi-Resolution Elastic Contrast Shapelets For Robust Arrhythmia Classification Using Extreme Gradient Boosting

Arrhythmia classification in continuous electrocardiogram (ECG) recordings remains a challenging task due to inter-patient variability, signal noise, and the non-stationary nature of heart rate variability. Traditional machine learning approaches heavily rely on rigid fiducial point extraction or fixed-length Euclidean similarity measures, which frequently fail under physiological time-warping. In this paper, we propose a robust feature-extraction and classification framework utilizing Multi-Resolution Elastic Contrast Shapelets combined with an Extreme Gradient Boosting (XGBoost) classifier. Our methodology addresses the need for rigid feature engineering by extracting variable-length, discriminative morphological primitives (shapelets) at both micro (beat) and macro (rhythm) resolutions. We evaluate these primitives using the FastDTW algorithm to generate a tabular pattern-based embedding. To ensure strict clinical validity and prevent intra-patient data leakage, the architecture is evaluated under a rigorous cross-database protocol using the harmonized MIT-BIH and Icentia11k datasets. Experimental results demonstrate that the proposed multi-resolution elastic approach significantly outperforms traditional rigid baselines, achieving an optimized global accuracy of 92.50%. By successfully mitigating "window myopia", the model yields highly balanced F1-Scores exceeding 0.90 across complex macro-morphologies (Atrial Fibrillation and Normal Sinus Rhythm) and 0.95 for localized micro-morphologies (Premature Atrial Contractions). This framework bridges the gap between state-of-the-art predictive performance and transparent clinical interpretability.