MSA-TCN: An Ultra-Compact, Energy-Efficient Edge TCN for Explainable HAR in Consumer IoHT Devices

Human Activity Recognition (HAR) in consumer healthcare and Internet of Health Things (IoHT) devices demands models that are accurate, energy-efficient, and explainable for on-device deployment. We propose MSA-TCN, an ultra-compact Adaptive Temporal Convolutional Network that combines multi-scale temporal convolutions, causal dilated encoding, and hybrid channel–temporal attention to model both short- and long-range motion dependencies. A subject-aware augmentation strategy enhances generalization across users and sensor configurations without additional computation. The quantized INT8 variant achieves an average accuracy of 98.7% across five benchmark datasets, with a compact size of 0.08 MB and an inference latency of 1.8 ms per 256-sample window on a mid-range smartphone (Snapdragon 845). Energy consumption is reduced by over 98% (from 32.5 mJ to 0.36 mJ) compared to FP32 inference. Post-hoc SHAP analysis provides clinically meaningful attributions, improving transparency and user trust in healthcare settings. By enabling on-device prediction, MSA-TCN minimizes wireless transmission and ensures reliability under intermittent connectivity, offering an interpretable, energy-efficient solution for next-generation IoHT applications.