| Abstract [eng] |
Atrial fibrillation (AFib) is a major risk factor for stroke, but it may be not detected in time due to being asymptomatic. Wearable devices and automatic AFib detection algorithms enable continuous monitoring and earlier detection. Numerous deep learning models have been proposed, but direct comparison between them is difficult due to different experimental protocols used in studies. Also their robustness to noise and use of diffusion-based denoising methods remain underexplored. Therefore, we evaluated three deep learning architectures: CTRhythm, MFEGNet, and MGCNet under clean and noisy conditions and tested the effect of diffusion-based denoiser DeScoD-ECG effect on AFib classification accuracy. These classifiers were trained on SHDB-AF dataset and cross tested on AFDB. CTRhythm, achieved the highest classification performance. Further experiments with different window lengths and loss functions identified 10-second windows and focal loss as optimal for model training, which resulted in macro F1 score 0.875 on external dataset. Noise robustness experiments revealed severe performance degradation across all noise types. Diffusion-based denoising recovered model performance for BW noise, and restored macro F1 from 0.561 to 0.954 on SHDB-AF dataset. However, denoising was limited for MA and EM noise. These findings highlight that classification performance on clean signals does not guarantee results on noisy signals, and that noise evaluation and robust denoising steps are an important part of any AFib detection pipelines. |
