Reclassifying Single Lead Sinus Rhythm ECGs for Improved AF Recurrence Detection With Deep Learning

Atrial fibrillation (AF) remains a common outcome after catheter ablation, with recurrence observed in up to 50% of patients within 5 years. Despite extensive research, challenges remain in optimizing observation and prediction windows for AF recurrence assessment. Therefore, we developed and validated a novel deep learning model that reclassifies sinus rhythm based on past and future AF events using single-lead ECG strips from handheld devices.

Of the 558 participants, 308 with no documented recurrence during study follow-up contributed Benign Sinus strips, while 250 with documented recurrence contributed sinus rhythm strips recorded within ±3 days of an AF event, defined as Severe Sinus strips. Both clinical feature-based models and a deep learning model were used to classify these ECGs according to their proximity to AF recurrence events. Performance was evaluated using area under the receiver operating characteristic curve, accuracy, sensitivity, specificity, and F1-score.

The strip-level pretrained deep learning model achieved an area under the receiver operating characteristic curve of 0.86, with sensitivity of 0.79, and specificity of 0.76. The complementary participant-level clinical feature-based Random Forest benchmark achieved an area under the receiver operating characteristic curve of 0.71, sensitivity of 0.58, and specificity of 0.69. Feature-based analyses identified associations involving heart-rate variability, mean R-R interval, female sex, left atrial volume, and mean P-wave duration.

Our dynamic deep learning model reclassifies 30-second single-lead sinus rhythm ECGs based on AF events within ±3 days, transforming a brief ECG snapshot into a summarized indicator of short-term AF risk.