A novel 1D generative adversarial network-based framework for atrial fibrillation detection using restored wrist photoplethysmography signals

Atrial fibrillation (AF) poses an increased stroke risk, necessitating effective detection methods. While electrocardiogram (ECG) is conventionally used for AF detection, the simplicity and suitability for long-term monitoring make photoplethysmography (PPG) an attractive alternative. In this study, we present a novel approach for AF detection utilizing smartwatch-based wrist PPG signals. Notably, this is the pioneering use of 1D CycleGAN (Generative Adversarial Network) for reconstructing 1D wrist PPG signals, addressing the challenges posed by poor signal quality due to motion artifacts and limitations in acquisition sites. The proposed method underwent validation on a dataset comprising 21,278 10 s long wrist PPG segments. Two experiments were conducted to evaluate 1D Self-AFNet's robustness by training on one split and testing on the other. First, the model was trained with Test Split 1 and evaluated on Test Split 2, then vice versa. Our classification model, Self-AFNet, incorporating 1D-CycleGAN restoration, demonstrated accuracy at 96.41 % and 97.09 % for the two splits, respectively. The restored signals exhibited a significant accuracy improvement (2.94 % and 5.08 % for test splits, respectively) compared to unrestored PPG. Additionally, AF detection using ECG signals, paired with matched PPG signals, confirmed the validity of employing reconstructed PPG for classification. Self-AFNet achieved impressive accuracies of 98.07 % and 98.97 %, mirroring the performance of AF detection using reconstructed PPG segments. This study establishes that reconstructed wrist PPG signals from wearable devices offer a reliable means for AF detection, contributing significantly to stroke risk reduction.