Schematic overview of study design and analyses. The AI-ECG age prediction model was trained on up to five electrocardiograms for each participant in the training split of the Severance dataset to predict their chronological age. The participants for training include those designated for validation during the training process. The artificial intelligence model was then validated in one hold-out test set and five external test sets. Participants lacking electrocardiograms with proper waveforms (12 leads/500 Hz/10 s), those aged below 20 or above 90, or without age information were excluded from the model development and validation phase (Exclusion 1). For the observational analysis phase on atrial fibrillation risk, we excluded participants who had a history of atrial fibrillation, a follow-up duration of less than a day, or an AF occurrence during the 30-day blanking period (Exclusion 2). Those without PRS-AF data were further excluded from the genetic interaction analysis phase (Exclusion 3). AF, atrial fibrillation; AI, artificial intelligence; ECG, electrocardiogram; ECG aging, electrocardiographic aging; N, numbers of participants; PRS-AF, polygenic risk score for atrial fibrillation; pt., participants; SHC, Severance Health Check-up