KAN-Former: a lightweight ECG model for real-time atrial fibrillation detection on wearable devices

Atrial fibrillation (AF) is the most prevalent sustained cardiac arrhythmia and a major contributor to stroke and heart failure. Continuous monitoring using wearable electrocardiogram (ECG) devices enables early detection, but deploying deep learning models on resource-constrained platforms remains challenging due to limited computation and stringent real-time latency requirements. To address these challenges, we propose KAN-Former, a lightweight hybrid ECG classification model. KAN-Former consists of: (1) a multi-scale convolutional front-end enhanced with channel attention to capture morphological details across different temporal resolutions; (2) a Nyström-approximated Transformer module that models global temporal dependencies with linear complexity; (3) a time-frequency feed-forward network that refines spectro-temporal rhythm representations; and (4) a sparsity-optimized residual KAN classifier that improves representational efficiency and reduces parameter count. Experimental results on PhysioNet 2017 and CPSC 2018 demonstrate that KAN-Former achieves accuracies of 94.5% and 94.6%, respectively, with only 1.16M parameters and an inference latency of 3.8 ms, outperforming state-of-the-art lightweight models while maintaining real-time on-device performance. Furthermore, even under aggressive pruning and INT8 quantization, the accuracy degradation remains below 0.5%, confirming the scalability of KAN-Former for wearable AF screening.