Ots-net: unlocking mechanistic interpretability in ECG arrhythmia classification

Deep learning models have achieved remarkable accuracy in electrocardiogram (ECG) arrhythmia classification, yet their “black-box” nature remains a major barrier to clinical trust. This paper introduces the Oscillatory Temporal Synchronization Net (OTS-Net), a novel, biophysically-plausible architecture designed to address this performance-interpretability deadlock. OTS-Net integrates a spatio-temporal convolutional backbone with Artificial Kuramoto Oscillatory Neurons (AKOrN), shifting the paradigm from static pattern recognition to dynamic synchronization modeling. By emulating the collective behavior of cardiac pacemaker cells, our approach achieves competitive performance −99.09% accuracy on the MIT-BIH Arrhythmia Database with an ultra-efficient 103 K parameters-while offering a framework for unprecedented mechanistic interpretability. We present an analytical suite that decodes the model’s internal state, visualizing its learned “anatomy” (coupling matrix) and emergent “physiology” (class-specific synchronization patterns). This enables diagnostic explanations based on quantifiable disruptions in phase-locking and network coherence, rather than opaque feature correlations. This work demonstrates that by grounding AI in the fundamental principles of biological synchronization, it is possible to create systems that are not only accurate and efficient, but are also inherently transparent and trustworthy.