Artificial intelligence (AI)–enhanced electrocardiography has been developed to detect paroxysmal atrial fibrillation (AF) using sinus rhythm ECGs (SR-ECGs). In this retrospective study published in the Annals of Noninvasive Electrocardiology, investigators developed and evaluated AI models using data from Tokai University (n=172,613; Nihon Kohden system) and The Cardiovascular Institute (n=19,170; GE MUSE system). AF-labeled SR-ECGs were defined as recordings obtained within 31 days of an AF episode, while SR-labeled ECGs required ≥1095 days of AF-free follow-up.

Three datasets were constructed, and five models were developed: scratch models (S1–S3) trained separately on individual datasets and fine-tuned models (F1, F2) pretrained on Dataset 3 and fine-tuned on Datasets 1 and 2. Model F2, fine-tuned on homogeneous cardiology department datasets, achieved AUCs of 0.885 (A1), 0.829 (A2), and 0.845 (A3). Model F1, fine-tuned on heterogeneous datasets, showed lower performance (AUCs 0.837, 0.726, and 0.660). Performance remained consistent across ECG format variants (B1–B3).

Limitations include retrospective design, data from two Japanese institutions only, lack of prospective validation, and potential influence of class imbalance despite mitigation strategies.

Fine-tuning on homogeneous datasets was associated with higher and more consistent model performance. Prospective validation in broader clinical settings is required to clarify real-world applicability.