A deep learning approach to monitoring and detecting atrial fibrillation using wearable technology

Atrial Fibrillation (AF) is the most common cardiac arrhythmia in clinical practice, with a prevalence of 2% in the community. Not only it is associated with reduced quality of life, but also increased risk of stroke and myocardial infarction. Unfortunately, many cases of AF are clinically silent and undiagnosed, but long-term monitoring is difficult. Nonetheless, efforts at monitoring at-risk individuals and detecting clinically silent AF may yield significant public health benefit, as individuals with new-onset, asymptomatic AF would receive preventive therapies with anticoagulants and beta-blockers, for example. Wearables have enormous potential to provide low-risk and low-cost long-term monitoring of AF, but signals from such devices suffer from significant movement related noise that resembles AF. This work presents a robust approach to AF detection in a sequence of short windows with significant movement artifact. Pulsatile photoplethysmographic (PPG) data and triaxial accelerometry from 98 subjects (45 with AF and 53 with other rhythms) were captured using a multichannel wrist-worn device. A single channel electrocardiogram (ECG) was recorded (for rhythm verification only) simultaneously. A novel deep neural network approach to classify AF from wrist-worn PPG signals was developed on this data. A continuous wavelet transform was applied to the PPG data and a convolutional neural network (CNN) was trained on the derived spectrograms to detect AF. Combining the output of the CNN with features calculated based on beat-to-beat variability and signal quality provided a significant accuracy boost. Leave-one-out cross validation resulted in a pooled AUC of 0.95 (Accuracy=91.8%). The proposed approach resulted in a novel robust and accurate algorithm for detection of AF from PPG data, which is scalable and likely to improve in accuracy as the dataset size continues to expand.