A lightweight residual dilated temporal transformer block for ECG classification on edge devices.

Wearable devices play a crucial role in healthcare by enabling continuous monitoring of vital physiological signals such as ECG, heart rate, respiration, body temperature, oxygen saturation, blood pressure, and activity metrics. These systems facilitate early detection and ongoing management of cardiovascular diseases, extending clinical capabilities beyond traditional settings. However, since the wearables are continually powered, there are stringent constraints on energy efficiency, processing latency, and data privacy. While cloud-based inference introduces high communication overhead, on-device deep learning raises computational and thermal challenges, demanding lightweight, accurate, and privacy-aware solutions. In this work, we propose a highly efficient and resource conscious deep learning model for ECG classification of three clinically significant classes: Arrhythmia (ARR), Congestive Heart Failure (CHF), and Normal Sinus Rhythm (NSR). The model is specifically optimized for deployment on edge devices. It combines residual dilated convolutions with a deep Transformer architecture enhanced by positional encoding, channel-wise attention (SE blocks), and class balanced data augmentation using SMOTE and jitter noise. Despite having only 692K parameters (2.64 MB) and approximately 0.234 GFLOPs, the model achieves 99.34% classification accuracy, Macro AUROC of 0.9996, Cohen's Kappa of 0.9891, and Log Loss of 0.0495. It further demonstrates strong generalization with a Hamming Loss of 0.0062 and Matthews Correlation Coefficient (MCC) of 0.9891 on the benchmark ECG dataset. The proposed method achieves an exceptional trade off between accuracy, model complexity, and inference speed, making it well suited for real-time, privacy preserving cardiac monitoring on low power, latency constrained platforms such as wireless body sensor networks and wearable edge devices.