FOUND-AF: Benchmarking ECG Foundation Models for Atrial Fibrillation Detection

A new benchmarking framework named FOUND-AF has been launched to assess how well ECG foundation models identify atrial fibrillation (AF), the most prevalent sustained cardiac arrhythmia linked to higher risks of stroke, heart failure, and death. This framework tackles the issue of inconsistent evaluations by offering a cohesive, leakage-controlled, and deployment-focused benchmark. Nine foundation models from five categories—HuBERT-ECG, CLEF, ST-MEM, ECG-JEPA, and ECGFounder—were evaluated under uniform experimental conditions using four diverse ECG datasets: AFDB, CinC2017, CPSC2021, and LTAFDB. All models served as frozen feature extractors with standardized preprocessing and model-specific resampling. The study, published on arXiv with ID 2608.03597, seeks to clarify the effectiveness of these models, which has been obscured by varying datasets and methodologies in prior research. The results aim to aid in choosing foundation models for automated AF detection in both clinical and research environments.