Sequential AI-ECG Diagnostic Protocol for Opportunistic Atrial Fibrillation Screening: A Retrospective Single-Center Study.

Abstract

Background/Objectives: Atrial fibrillation (AF) often occurs in episodes that are sudden and go unnoticed, reducing the chances of anticoagulation. We evaluated a two-stage AI ECG screening protocol that uses a single ECG model at initial screening and, if necessary, a serial ECG model after short interval follow-up to enhance accuracy while saving monitoring resources. Methods: We analyzed 248,612 12-lead ECGs from 164,793 adults (AF, n = 10,735) for model development and assessed the protocol in 11,349 eligible patients with longitudinal ECGs. The proposed algorithm first applied a single-ECG AI model at the initial visit, followed by a serial-ECG AI model three months later if AF was not initially detected. The model's performance was evaluated using several metrics, including the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, accuracy, and F1 score. Results: The protocol achieved an AUROC of 0.908 with a sensitivity of 88.1%, specificity of 78.7%, positive predictive value (PPV) of 30.2%, negative predictive value (NPV) of 98.4%, accuracy of 79.6%, and an F1 score of 0.450. Among patients with a history of stroke (n = 551), 84.9% were correctly identified as AF-positive under the protocol. Conclusions: A sequential AI ECG strategy maintains high NPV at entry and improves PPV with longitudinal confirmation. This approach can prioritize ambulatory monitoring for those most likely to benefit and merits prospective, multi-center validation and cost-effectiveness assessment.