Abstract
OBJECTIVE: Smartwatches with photoplethysmographic (PPG) sensors are ideal for early atrial fibrillation (AF) detection through continuous monitoring. However, prior deep learning was limited either to controlled environments, to minimize motion artifacts, or to short duration data collection. Additionally, premature atrial/ventricular contractions (PAC/PVC), which often confound AF detection algorithms, remains understudied due to limited datasets. Current state-of-the-art methods achieve only 75% sensitivity for PAC/PVC class on minimally motion artifact corrupted PPG data, despite showing 97% AF detection accuracy. METHODS: We addressed the above limitations using data from the recently completed NIH-funded Pulsewatch clinical trial which collected over two weeks of smartwatch PPG data from 106 subjects. Our computationally efficient 1D bi-directional Gated Recurrent Unit deep learning model incorporated multi-modal inputs (1D PPG, accelerometer, and heart rate data) to classify normal sinus rhythm, AF, and PAC/PVC. RESULTS: Our model achieved an unprecedented 83% sensitivity for PAC/PVC detection while maintaining a high accuracy of 97.31% for AF detection, outperforming the best retrained state-of-the-art model by 20.81% and 2.55%, respectively. It was also 14 times more computationally efficient and 2.7 times faster. Testing on two external PPG datasets collected with a different smartwatch and a fingertip PPG sensor, our model demonstrated better generalizability with macro-averaged AUROC values of 96.22% and 94.17%, respectively. CONCLUSION: A light-weight multimodal input deep learning model can accurately distinguish PAC/PVC from AF, reducing false positive detection of AF. SIGNIFICANCE: Accurate AF and PAC/PVC detection with minimal false positive detection can enhance clinical and public acceptance of smartwatch-based AF monitoring.