Application of Artificial Intelligence in the Detection and Prediction of Atrial Fibrillation in ECG Recordings – A Narrative Review

Introduction and purpose. Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia and a leading cause of ischaemic stroke, yet it frequently remains undiagnosed because of its paroxysmal and often asymptomatic nature. Artificial intelligence (AI), and in particular deep learning applied to the electrocardiogram (ECG), has emerged as a transformative approach both for detecting AF when the arrhythmia is present and for predicting future AF from an ECG recorded during sinus rhythm. The purpose of this narrative review is to summarise the current state of knowledge on AI-enabled ECG analysis for the detection and prediction of AF and to critically appraise its clinical readiness. Description of the state of knowledge. Convolutional neural networks can classify AF on single-lead and twelve-lead ECG with cardiologist-level accuracy (area under the curve, AUC, above 0.97 for rhythm classification) and can identify an "electrocardiographic signature" of AF on a normal sinus-rhythm tracing (AUC approximately 0.87). Consumer wearables using photoplethysmography and single-lead ECG achieve pooled sensitivity and specificity around 95%. Predictive models for incident AF reach AUC values of 0.78–0.85 and have been validated prospectively, enriching the diagnostic yield of screening. Nevertheless, false positives, unclassifiable tracings, limited external validity, and the absence of trials demonstrating a reduction in stroke temper enthusiasm. Summary. AI-ECG is a mature and accurate tool for AF detection and a promising instrument for AF prediction, but its translation into improved patient-relevant outcomes remains unproven. Rigorous external validation, outcome-driven randomised trials, and clear regulatory and implementation frameworks are required before routine clinical adoption.