Device-aware compensation of nonideal spintronic synapses for electrocardiogram classification using spiking neural networks

Abstract

Spintronic synapses offer a promising route toward low-power neuromorphic hardware for edge intelligence. Their nonvolatility, multilevel tunability, and compatibility with event-driven spiking neural networks (SNNs) make them attractive for temporal biomedical signal processing. Moving beyond conventional demonstrations of synaptic plasticity, here we examine how nonideal spin–orbit torque (SOT) synaptic updates affect SNN learning and how they can be compensated. TaMo/CoFeB/MgO/Ta SOT devices with perpendicular magnetic anisotropy exhibit nonvolatile multilevel anomalous Hall resistance states and reversible long-term potentiation and depression. Combined finite-element and micromagnetic simulations indicate that the nonlinear weight evolution is closely associated with spatially nonuniform SOT-driven domain dynamics. Incorporating the measured update curves into an SNN for MIT-BIH electrocardiogram abnormal-beat classification shows that device nonideality strongly degrades recognition of minority-class abnormal beats. A device-aware compensation strategy restores balanced effective updates, bringing the overall accuracy close to the ideal-update reference (95.06% vs 95.23%). It also increases the F1 score for the minority F-class abnormal beats from 11.86% to 64.08%. This work highlights device-aware compensation as an effective route for translating nonideal spintronic synapses into practical low-power biomedical neuromorphic hardware. Such hardware is promising for wearable health-monitoring applications.