Leveraging ECG foundation models in critical care for sinus rhythm and atrial fibrillation classification

Authors: Maria Galanty, Michiel Hulleman, Merijn C. Reuland, Björn van der Ster, Alexander P. J. Vlaar, Clara I. Sánchez & the AF Expert Panel

This study evaluated ECG foundation models for classifying sinus rhythm and atrial fibrillation from continuous ICU ECG monitoring. The investigators assessed an existing pretrained model on expertly annotated ICU ECG data and subsequently evaluated targeted fine-tuning using selected public datasets, with and without additional weakly labelled ICU data.

The off-the-shelf model showed limited transferability to ICU recordings (F1 0.68 for sinus rhythm and 0.52 for atrial fibrillation). Fine-tuning improved performance (F1 0.91 and 0.82, respectively), while adding weakly labelled ICU data increased recall with some reduction in precision. The results highlight the importance of dataset quality, targeted fine-tuning and local validation.