Abstract
We introduce a lightweight 1D ConvNeXtV2-based neural network for the robust detection of atrial fibrillation (AFib) and atrial flutter (AFL) from single-lead ECG signals. Trained on multiple public datasets (Icentia11k, CPSC-2018/2021, LTAF, PTB-XL, PCC-2017) and evaluated on MIT-AFDB, MIT-ADB, and NST, our model attained a state-of-the-art F1-score of 0.986 on MIT-AFDB. With only 770 k parameters and 46 MFLOPs per 10 s window, the network remained computationally efficient. Guided Grad-CAM visualizations confirmed attention to clinically relevant P-wave morphology and R-R interval irregularities. This interpretable architecture is, therefore, well-suited for deployment in resource-constrained wearable or bedside monitors. Future work will extend this framework to multi-lead ECGs and a broader spectrum of arrhythmias.