Background: Short-term prediction of Atrial Fibrillation (AF) onset from sinus-rhythm wearable ECG could enable early warning. Several algorithms have already been proposed, notably on the PhysioNet AFPDB, a small benchmark of 200 30-minute ECG records, where published studies reported optimistic performance above 90%; however, independent reproducibility analyses showed that these results could not be reliably reproduced, highlighting the need for larger, better-controlled datasets.
Aim: To compare the short-term AF onset prediction performance of state-of-the art machine learning and deep learning models using a single large-scale reference dataset.
Methods: We used the IRIDIA database with manually annotated AF onsets. Positive samples were 30-minute sinus-rhythm segments ending 5 minutes before AF onset. A total of 1,347 AF episodes were identified across 570 patients, and 783 episodes from 494 patients met the forecasting inclusion criteria. Patient-level train/validation/test splits were used. We compared a baseline XGBoost model using five selected Heart Rate Variability (HRV) features (RMSSD, PAS, SODP Q1, SODP CTM100, SD1/SD2), and deep learning architectures trained on combinations of HRV features, raw ECG, and RR-interval sequences.
Results: The HRV XGBoost baseline achieved an AUROC of 0.657. An ECG-only transformer performed worse (AUROC 0.587), while a hybrid ECG+HRV transformer reached AUROC 0.642. RR-based models remained competitive (RR-only AUROC 0.664; hybrid RR+HRV AUROC 0.668). The best overall discrimination was obtained with a hybrid pretrained foundation model (CLEF-Small+HRV: AUROC 0.680). Increasing backbone size to CLEF-Medium did not improve performance (AUROC 0.666).
Conclusions: In this setting, the most effective strategy is to use hybrid approaches including HRV in the input features. Transfer learning based on a foundation ECG model enhanced with selected HRV features yields the best performance overall.