Introduction: Atrial fibrillation (AF), the most prevalent sustained cardiac arrhythmia, affects roughly 2% of the general population and is linked to stroke, heart failure, and higher mortality. AF may occur in brief episodes that are often missed on a standard resting 12-lead, 10-second electrocardiogram (ECG), necessitating longer recordings. However, these extended recordings are impractical for manual review by clinicians, creating a need for computerized, automated detection of atrial fibrillation.
Purpose: This study aims to extend state-of-the-art research based on the IRIDIA-AF database by addressing open questions about using both 2D spectrograms and 1D ECG window segments as inputs to convolutional neural network (CNN) models for AF prediction and detection.
Methodology: The IRIDIA-AF database contains 167 records from 152 patients with paroxysmal AF. Each record lasts 19 to 95 hours and is split into 24-hour files. To forecast and detect AF, 10- and 30-second 1D ECG windows and 2D spectrograms from leads I and II were generated. For AF forecasting, 355 pre-AF and normal case pairs (710 segments) were used, whereas AF detection employed 4500 AF and normal case pairs (9000 segments).
Results: As shown in Table 1, for forecasting tasks, models based on 2D spectrograms generally outperformed those using 1D raw ECG data, whereas for detection, the two approaches yielded comparable performance.
Conclusion: Although detection performance did not differ substantially between the 1D and 2D models, there was a marked difference in forecasting performance. The 2D model, in particular, achieved notably strong forecasting results, approaching the performance level of the detection models.