AFLoc8-Net: End-to-End Atrial Fibrillation Localization with a Dual Head Transformer.

Paroxysmal Atrial Fibrillation (AFib) poses significant health risks, making its timely and precise detection crucial. Traditional methods for AFib detection in ambulatory recordings often rely on separate classification and post-processing steps, which can lead to inaccuracies in identifying the precise timing and duration of AFib episodes. To address these limitations, we introduce 'AFLoc8-Net' (Atrial Fibrillation Locator Network), a novel end-to-end transformer-based model designed for the robust detection and localization of AFib episodes in long-term ECG recordings. Our model integrates both classification and regression heads within a transformer framework, allowing it to simultaneously classify AFib events and accurately determine their onset and duration. We evaluate our approach with the winners of the China Physiological Signal Challenge (CPSC) 2021 on a separate test dataset. We show the drawbacks of an N-beat classification method and how our approach mitigates these challenges.