Electrophysiology-Guided Channel Pruning for Rare Arrhythmia Detection in Quantized Edge Neural Networks

Abstract

We propose a domain-aware channel pruning framework for quantized 1D convolutional neural networks deployed on low-power microcontrollers for real-time arrhythmia detection in wearable ECG patches. Conventional pruning methods rely solely on activation magnitude statistics, which often discard channels critical for detecting rare but clinically significant arrhythmia morphologies such as fragmented QRS complexes, epsilon waves, and micro-volt T-wave alternans. To address this limitation, we introduce an electrophysiological sensitivity score that integrates activation magnitude with cosine similarity between learned convolutional filters and canonical template filters derived from cardiovascular electrophysiology models. These templates are constructed as damped sinusoids, periodic amplitude modulations, and multi-peak Gaussian mixtures that capture the morphological signatures of rare arrhythmia markers. The pruning module ranks channels by this combined score and removes those with the lowest scores, thereby preserving clinically relevant features. The remaining weights are re-quantized to 8-bit integers using min-max quantization, and the network is fine-tuned with class-balanced mini-batches and a weighted cross-entropy loss that over-samples rare arrhythmia episodes. Our method is evaluated on a 5-layer 1D-CNN with 5 output classes, processing 5-second ECG segments at 200 Hz. On an STM32L4 microcontroller, the pruned model achieves a 2.3x speedup in inference time and a 1.8x reduction in energy per inference compared to the unpruned baseline. Furthermore, it maintains a sensitivity of 94.2% for rare arrhythmia classes, significantly outperforming uniform magnitude-based pruning which yields only 88.1% sensitivity. The primary contribution is a principled integration of electrophysiological priors into the neural network compression pipeline, enabling edge devices to detect rare cardiac events without sacrificing computational efficiency.