Prune or Quantize? Layer-wise Compression of Time-Series ECG Foundation Networks

Foundation models for biosignals, such as wearable ECG monitors, face challenges in resource-constrained settings due to high memory and computational demands. We propose an adaptive layer-wise compression framework that combines quantization and pruning to reduce model size while preserving predictive performance. Layer importance, estimated via parameter contribution and weight variance, guides fine-grained assignment of bit-widths and pruning thresholds, balancing efficiency and accuracy across high- and low-sensitivity layers. Experiments on Chapman and CPSC ECG datasets show that our method consistently outperforms fixed global compression schemes, achieving up to 10.44$\times$ compression with no loss in performance. Our architecture-agnostic framework scales from lightweight residual networks to large foundation models, enabling real-time, low-resource ECG monitoring. By efficiently deploying foundation models on edge devices, this work advances scalable, physiology-aware biosignal AI for mobile health and clinical applications.