Shinya Suzuki, Mari Amino, Nobumoto Moriai, Yoshifumi SUMIOKA, Akira Hyodo, Keiichi Satoh, Tsuneo Takayanagi, Takuya Umemoto, Yota Sekiguchi, Naomi Hirota, Atsuhiko Yagishita, Yuji Ikari, Taira Nakayama, Yoichi Ohnuki, Eiichiro Nagata, Jun Motogi, Takeshi Yamashita, Koichiro Yoshioka
BACKGROUND: Artificial intelligence (AI)-enhanced electrocardiography (ECG) has been developed to detect paroxysmal atrial fibrillation (AF) from sinus rhythm ECGs (SR-ECGs). For broader applicability, model development across institutions and ECG systems is essential. METHODS: We developed an AI-enhanced ECG model using SR-ECGs from Tokai University (n = 172,613; Nihon Kohden [NK] system) and The Cardiovascular Institute (n = 19,170; GE MUSE system). AF-labeled SR-ECGs were defined as recordings within 31 days of an AF episode, while SR-labeled SR-ECGs were those with ≥ 1095 days of AF-free follow-up. Three datasets were constructed: Dataset 1 (Tokai University, all departments, NK), Dataset 2 (Tokai University, Cardiology Department, NK), and Dataset 3 (The Cardiovascular Institute, Cardiology Department, MUSE). We developed five models: scratch models (S1-S3) trained on Datasets 1-3, and fine-tuned models (F1, F2) trained on Datasets 1 and 2 after pretraining on Dataset 3. Models were evaluated using A1-A3 (same as Datasets 1-3) and B1-B3, which differed in ECG resolution and compression (B1: original MUSE, B2: MUSE-NK intermediate, B3: NK-converted). RESULTS: Model F2, fine-tuned on homogeneous datasets from cardiology departments, showed consistently high performance (AUC: A1 = 0.885, A2 = 0.829, A3 = 0.845). Model F1, fine-tuned on heterogeneous datasets, demonstrated lower performance (AUC: A1 = 0.837, A2 = 0.726, A3 = 0.660). Model performance was consistent across different ECG format variants (B1-B3). CONCLUSION: Fine-tuning on homogeneous data improved performance and generalizability, whereas heterogeneous data led to reduced performance. ECG system format differences had minimal impact on model accuracy.