Artificial Intelligence Applied to Electrocardiograms Recorded in Sinus Rhythm for Detection and Prediction of Atrial...

Background and Objectives: Subclinical paroxysmal atrial fibrillation (AF) is often undetected by conventional screening strategies, until complications emerge. Artificial intelligence (AI) applied to sinus rhythm electrocardiograms has emerged as a promising tool to identify individuals with occult AF and to predict the risk of future incident AF. This scoping review synthesizes evidence from original studies evaluating AI models trained on sinus rhythm ECGs for AF detection or AF prediction. Materials and Methods: A comprehensive search of MEDLINE, Embase, Web of Science, Scopus, and IEEE Xplore was conducted to identify peer-reviewed studies from inception to November 2025. Eligible studies included original investigations in which the model input was a sinus rhythm ECG and the outcome was either paroxysmal AF or new-onset AF. Extracted variables included cohort characteristics, ECG acquisition parameters, AI architecture, model predictive performance, AF prediction horizon, clinical outcomes, and validation strategy. Risk of bias was assessed using PROBAST. Results: Nineteen studies met the inclusion criteria. Retrospective datasets ranging from several thousand to over one million ECGs and convolutional or deep neural network AI architectures were used in most studies. AI-ECG models demonstrated high diagnostic accuracy for detecting subclinical AF (ten studies; AUROC 0.75-0.90) and for predicting long-term new-onset AF (six studies; AUROC 0.69-0.85) from a single sinus rhythm ECG. Robust external validation was reported in eleven studies. Combining AI-ECG models with clinical risk factors improved AF predictive performance in several reports. Key limitations across studies included retrospective design, patient selection, limited calibration reporting, and sparse prospective impact data. Conclusions: AI-based analysis of sinus rhythm ECGs can detect occult AF and stratify future AF risk with moderate-to-high accuracy across multiple populations and healthcare systems. However, rigorous prospective trials, evaluating clinical benefit, cost-effectiveness, calibration across demographic groups, and real-world implementation, are required before broad adoption in clinical practice.