Aim: While atrial fibrillation (AF) is the most common arrhythmia, it is rarely detected in routine clinical electrocardiograms (ECGs). Previous stud-ies have suggested that patients with AF can be distinguished from those without AF based on sinus rhythm electrocardiograms using machine learn-ing. Despite this, it remains uncertain how well such models generalize across datasets acquired from different Holter devices. In this study, we in-vestigated the ability to distinguish individuals with and without paroxys-mal AF within a 24-hour window using heart rate variability (HRV) features derived from beat-to-beat (RR) intervals.
Methods: Two independent datasets were utilized, collected with distinct 2-lead Holter systems, differing in original recording duration, country of origin and lead positioning. From each 24-hour window, multiple RR seg-ments of 500 consecutive beats were extracted at least one hour away from AF episodes. A total of 27 HRV features were computed spanning time, fre-quency, non-linear, fragmentation, and recurrence plot-based domains. A random forest classifier was trained using nested cross-validation with greedy forward feature selection.
Results: When evaluated separately, the model trained on the smaller short-term Holter dataset achieved higher performance than the model trained on the larger long-term Holter dataset (ROC-AUC: 0.81 vs 0.75). However, during cross-dataset evaluation, the model trained on the larger dataset generalized better while the model trained on the smaller dataset showed reduced performance (ROC-AUC: 0.85 vs 0.67).
Discussion: These results highlight the potential of HRV-based approach-es, while demonstrating generalization across different Holter devices. This is supported by the consistent selection of HRV features, including SODP Q1 and SD1/SD2, among datasets and studies. Overall, these findings emphasize the importance of dataset diversity for robust cross-dataset generalization.