Single-lead ECG AI model with risk factors detects Atrial Fibrillation during Sinus Rhythm

Abstract

Background and Aims Guidelines recommend opportunistic screening for atrial fibrillation (AF), using a 30-second single-lead electrocardiogram (ECG) recorded by a wearable device. Since many patients have paroxysmal AF, identification of patients at high risk presenting in sinus rhythm may increase the yield of subsequent long-term cardiac monitoring. The aim is evaluating an AI-algorithm trained on 10-second single-lead ECG with or without risk factors to predict AF. Methods This retrospective study used 13479 ECGs from AF-patients in sinus rhythm around time of diagnosis and 53916 age- and sex-matched control ECGs, augmented with seventeen risk factors extracted from electronic health records. AI models were trained and compared using one- or twelve-lead ECGs, with or without risk factors. Model bias was evaluated by age- and sex-stratification of results. Random forest models identified the most relevant risk factors. Results The single-lead model achieved an AUC of 0.74, which increased to 0.76 by adding six risk factors (95% confidence interval: 0.74-0.79). This model matched the performance of a twelve-lead model. Results are stable for both sexes, over ages ranging from 40 to 90 years. Out of seventeen clinical variables, six were sufficient for optimal accuracy of the model: hypertension, heart failure, valvular disease, history of myocardial infarction, age and sex. Conclusions An AI model using a single-lead sinus rhythm ECG and six risk factors can identify patients with concurrent AF with similar accuracy as a 12-lead ECG-AI model. An age- and sex matched dataset leads to an unbiased model with consistent predictions across age groups.