Atrial fibrillation (AF) is a prevalent diagnosis in Western countries and is associated with significant morbidity and mortality if untreated. Due to its initial paroxysmal nature, long-term or regular short-term electrocardiograms (ECGs) are required for diagnosis of AF. Wearables with reduced leads can be used for this purpose, although they provide less information than standard 12-lead ECGs.
We examined the impact of reduced-lead ECGs on the accuracy of AF detection, using interpretable features extracted from 2,478 12-lead ECGs of 30 seconds each from 2 large public databases. Furthermore, we investigated the applicability of our approach prospectively to ECGs from mobile patches. In addition to recording-specific metadata, the feature set included 123 lead-dependent features from the following domains: template-based morphology, morphology variability, F wave, heart rate variability, and signal quality indices. We trained decision tree ensembles for 8 different lead combinations, ranging from single Einthoven leads to 12 leads, to differentiate between AF, normal sinus rhythm (NSR), and other arrhythmias using a one-versus-all approach. For model optimization, we conducted an extensive grid search with 7 factors and 2,304 factor combinations for each lead combination.
Although the discriminability of AF, NSR, and other arrhythmias decreased with the number of leads, we achieved at minimum an F1 score of 0.918 when using only lead III, compared to the highest F1 score of 0.961 when using 12-lead ECG, based on 5-fold cross-validation. The 11 most important features were template-based morphology or morphology variability features, primarily related to the P wave.
Our results demonstrate high classification quality with reduced leads, indicating great potential for wearables to detect anomalies such as AF. Using a 3-lead model, we were able to detect AF episodes in mobile long-term ECGs, proving the potential of our approach for mobile applications.