Deep learning–based detection of ectopic beats from photoplethysmography recordings: external validation across invasive electrophysiology cohorts

Background Frequent ectopic beats are associated with incident atrial fibrillation and heart failure. Photoplethysmography (PPG)–based wearable devices enable scalable and continuous heart rhythm monitoring and may facilitate early intervention [1-3]. However, no consumer-grade device currently quantifies ectopic burden. Purpose To develop and externally validate a deep learning framework capable of discriminating ectopic beats from sinus rhythm (SR) and ventricular tachycardia (VT) using short PPG recordings. Methods The model was developed using 19,168 eight-second PPG segments acquired simultaneously with ECG during invasive electrophysiological studies and cardiac surgery in 42 patients at a tertiary UK centre (class distribution: 42.99% SR, 24.07% ectopy [ECT = premature atrial contractions + premature ventricular contractions], 32.94% VT). External validation was performed using 118,217 PPG segments from 91 patients undergoing invasive electrophysiological studies at an independent Chinese centre (58.46% SR, 32.82% ECT, 8.72% VT) [4]. All ECG recordings were labelled by expert reviewers blinded to the PPG waveforms. The deep learning framework comprised signal preprocessing, augmentation, feature engineering, a convolutional encoder, and a shallow classifier. Fiducial masks were incorporated to emphasise physiologically meaningful pulse waveform regions, alongside a signal quality–aware attention mechanism to adaptively weight features according to PPG quality. Performance was compared with support vector machine, and random forest classifiers based on handcrafted morphological PPG features. Results On the internal test set (20% hold-out), the proposed model achieved a macro-average area under the curve (AUC) of 0.92, sensitivity 0.79, specificity 0.90, and precision 0.79. Without retraining, performance on the external dataset remained robust, with a macro-average AUC of 0.95, sensitivity 0.80, specificity 0.94, and precision 0.76. Detailed metrics are provided in the main figure. Post hoc analyses demonstrated a positive correlation between classification performance and signal quality. Stable monomorphic VT was more frequently misclassified as SR, whereas short, unstable, or polymorphic VT episodes were more commonly misclassified as ECT. Conclusion This deep learning framework accurately identifies ectopic beats from short PPG recordings and demonstrates strong external generalisability. These findings support the use of PPG-based wearable devices to enable continuous ectopic burden monitoring and justify prospective clinical evaluation.