Hybrid feature engineering for resource-efficient Arrhythmia detection in electrocardiogram signals: An interpretable, separability-driven framework.

Abstract

Deep neural networks dominate automated arrhythmia detection, yet their reported performance often relies on intra-patient evaluation, which allows models to exploit patient-specific morphology rather than learn disease-related patterns. Such shortcuts are unavailable in deployed edge monitors, motivating a fundamental question: is arrhythmia detection primarily limited by classifier capacity or by representation quality? We investigate this question under a leakage-controlled inter-patient protocol. Six feature families, including causal graph-based and heart-rate-variability descriptors computed only from preceding beats, are extracted using only the training partition to prevent leakage of patient identity or future information into the representation. On MIT-BIH, a linear support vector classifier reaches a weighted F1 score of 0.780 and a four-class macro F1 score of 0.493. On INCART, evaluated using disjoint patient groups, the SVC reaches a weighted F1 of 0.844, and logistic regression a weighted F1 of 0.847. Near-linear separability is measured directly. A radial-basis-function kernel improves on the linear boundary by only 0.007 accuracy under grouped cross-validation (95% CI: [-0.011,+0.026], paired-test p=0.34), indicating no measurable benefit from a non-linear decision boundary. Accordingly, the residual difficulty lies in the representation rather than in classifier capacity. A same-budget convolutional network evaluated under the identical protocol is less accurate overall and collapses the supraventricular class, demonstrating that representation quality, rather than additional classifier capacity, governs rare-beat recovery. The linear model recovers the frequency evidence the network attends to, with descriptive band-importance agreement (cosine 0.95-0.97, identical rank ordering, and a shared dominant band) between the linear coefficients and the CNN's gradient attributions, using about 43 times fewer parameters and a 2.17 KB footprint. Under leakage-controlled inter-patient evaluation, causal physiological representations enable an interpretable linear model to achieve efficient and competitive arrhythmia detection potentially suitable for resource-constrained wearable and bedside decision-support systems, matching the discriminative evidence exploited by a comparable deep network.