Abstract
Wearable health devices have moved from step-counting toys to genuine medical instruments, but the gap between what they can sense and what they can decide on-device is still wide. Most clinical-grade pattern recognition still happens in the cloud, which costs latency, privacy, and battery life. In this paper, we propose NeuroPulse-Edge, a lightweight on-device architecture that combines spiking neurons with a spike-only attention block to deliver real-time anomaly detection on wearable biosignals at a sub-2 mW power budget. The system encodes ECG, PPG, accelerometer, and skin-temperature streams into spike trains through a delta-modulation front end, runs them through a stack of leaky integrate-and-fire neurons, a spike convolution block, and a binary spiking-attention block, and ends in a small INT8 classifier. The attention block is the key piece — it preserves the long-range temporal sensitivity that arrhythmia recognition needs, but does it with logical AND-OR and popcount instead of floating-point multiply-accumulates, so it costs almost nothing to run on a microcontroller. We evaluate the system on four 2025–2026 wearable benchmarks (CACHET-CADB, WildPPG, the 2025 multimodal stress dataset, and a 2026 long-term smartwatch arrhythmia release) against five strong baselines. NeuroPulse-Edge reaches 92.1% accuracy and a 0.921 macro F1 on a five-class arrhythmia task, draws 1.17 mW continuously on an ARM Cortex-M4 testbench, and answers in 14 ms per inference — an 8.2× power reduction and a 4.4× latency reduction over the strongest deep-learning baseline at the same accuracy, with every gain confirmed at p < 0.01 (Wilcoxon signed-rank with Holm–Bonferroni correction). Embedded analyses establish a surrogate-gradient convergence bound and a spike-sparsity energy bound for the architecture. The findings indicate that spike-based attention is a practical building block for clinical-grade wearable AI.
Open public full textRights: ALLOWED · cc-by-sa · Source: journal-auto-sync:CROSSREF_ISSN