Artificial Intelligence for Prediction and Detection of Atrial Fibrillation from Sinus-Rhythm Electrocardiograms and Ambulatory Monitoring

Atrial fibrillation (AF) is a highly prevalent arrhythmia associated with stroke, heart failure and excess mortality. Yet, “silent” AF episodes remain undetected, leading to underestimation of disease burden. Additionally, paroxysms occur in an “unpredictable” way, and available clinical scores only stratify long-term AF risk with moderate discrimination, lacking the ability to evaluate near-term events. Artificial intelligence (AI) applied to sinus rhythm from short or continuous electrocardiogram (ECG) recordings shows that such predictive information is hidden in “plain sight.” This complementary approach seeks to uncover latent AF substrate and forecast imminent AF episodes. Deep-learning models trained on 10-s, 12-lead ECGs can identify individuals with prevalent or long- or near-term AF with areas under the curve (AUCs) up to 0.90, outperforming established clinical scores. Image-based AI-ECG models extend these capabilities to paper or scanned ECGs. Furthermore, AI algorithms applied to 24-h Holter and multi-day patch recordings achieve AUCs ≥0.80 for detecting occult AF or predicting it within 14 days, consistently surpassing risk scores like C2HEST and HATCH. Short-term models utilizing heart-rate variability features further demonstrate that AF can be anticipated minutes to hours before onset, with accuracies around 90% in curated datasets. However, most AI-AF studies remain retrospective, single-system and focused on diagnostic yield rather than clinical outcomes like stroke or mortality. Moreover, few pragmatic trials have evaluated AI-guided AF screening and its translation into clinical benefit. Robust prospective trials and standardized evaluation frameworks are needed before AI-guided AF prediction can be routinely integrated into clinical decision-making.