Atrial fibrillation detection on reconstructed photoplethysmography signals collected from a smartwatch using a denoising autoencoder

Photoplethysmography (PPG) signals collected by wearables have been shown to be effective in accurate detection of atrial fibrillation (AF), provided that the data are devoid of motion and noise artifacts (MNA). Many studies have been previously conducted to detect AF arrhythmia using PPG data; however, the subjects were mostly in clinics or controlled settings with data collection lasting several minutes to at most several hours with minimal MNA. Our study, Pulsewatch, differs from previous AF studies in that PPG data from smartwatches prescribed to stroke survivors were continuously collected for two weeks in real-life conditions, which invariably included a significant amount of MNA. Our aim is to provide a framework for a novel use of a denoising autoencoder to reconstruct motion-artifact-removed PPG signals so that we can improve the AF detection per-formance and to increase the amount of analyzable data.We used more than 30,000 25-sec PPG segments from 129 subjects randomly selected from Pulsewatch and Stanford University's datasets. The training and testing datasets from these two databases came from smart-watches from different vendors with varying sampling frequencies and time duration of recordings in diverse and realistic settings. In this study, the highly corrupted PPG data were automatically detected and discarded, but those segments contaminated with low-to-moderate motion and noise artifacts (MNA) were subjected to a convolutional denoising autoencoder (CDA). To reconstruct the artifact-removed PPG segments, we proposed to employ two distinct CDA models for AF and non-AF data groups initially classified as AF or non-AF. Using the proposed approach, we significantly improved the performance of detecting occult AF. We achieved classifica-tion accuracy, sensitivity, and specificity of 91.02%, 91.54%, and 90.85%, respectively, for out-of-sample test data from both databases. By sanitizing data from low-to-moderate MNA, we were able to increase the usable data coverage by 21%.