Rhythm-Adaptive Recognition of ECG Pathologies Using Model-Informed Preprocessing and Classical Machine Learning Classifiers
Abstract
Automatic ECG pathology recognition is increasingly dominated by deep neural networks trained on very large datasets. Although such approaches can be highly accurate, they often require extensive labeled corpora, considerable computational resources, and post hoc explainability tools. This paper investigates a lighter hybrid alternative for multiclass ECG pathology recognition in which the diagnostic representation is formed by rhythm-adaptive processing grounded in the cyclic random-process model of ECG signals, while the final decision is made by standard machine learning classifiers. The pipeline reuses the previously validated stages of ECG preprocessing, rhythm-function estimation, and cycle segmentation, and adapts them from biometric authentication [1] to pathology recognition. The experimental material contained 14 ECG recordings acquired from 14 different patients and consolidated into five pathology groups derived from diagnostic file-name markers: extrasystolic activity, atrial fibrillation, paired ectopic beats, frequent ventricular ectopy, and conduction block morphology. Seven classifiers were studied: k-nearest neighbours, Linear SVM, Gaussian Process, Decision Tree, Random Forest, multilayer perceptron, and AdaBoost. Two validation protocols were considered: a within-recording cycle-level split and a stricter patient-disjoint evaluation on the multi-patient classes. In this pilot study, the proposed rhythm-adaptive representation showed promising performance while preserving low complexity and interpretability. The best Protocol A result was achieved by the Gaussian Process classifier: 97.89% accuracy and 97.48% macro-F1. Since the dataset included only 14 patient recordings and two single-patient pathology groups, the results should be viewed as preliminary feasibility evidence, not a definitive estimate of clinical generalization. Further validation on larger multi-patient datasets and broader pathology coverage is required.