Abstract
Wearable devices enable theoretically continuous, longitudinal monitoring of physiological measurements such as step count, energy expenditure, and heart rate. Although the classification of abnormal cardiac rhythms such as atrial fibrillation from wearable devices has great potential, commercial algorithms remain proprietary and tend to focus on heart rate variability derived from green spectrum LED sensors placed on the wrist, where noise remains an unsolved problem. Here we develop DeepBeat, a multitask deep learning method to jointly assess signal quality and arrhythmia event detection in wearable photoplethysmography devices for real-time detection of atrial fibrillation. The model is trained on approximately one million simulated unlabeled physiological signals and fine-tuned on a curated dataset of over 500 K labeled signals from over 100 individuals from 3 different wearable devices. We demonstrate that, in comparison with a single-task model, our architecture using unsupervised transfer learning through convolutional denoising autoencoders dramatically improves the performance of atrial fibrillation detection from a F1 score of 0.54 to 0.96. We also include in our evaluation a prospectively derived replication cohort of ambulatory participants where the algorithm performed with high sensitivity (0.98), specificity (0.99), and F1 score (0.93). We show that two-stage training can help address the unbalanced data problem common to biomedical applications, where large-scale well-annotated datasets are hard to generate due to the expense of manual annotation, data acquisition, and participant privacy.
Highlights
Wearable devices are increasingly used in cardiology for out-ofthe-clinic healthcare monitoring[1]
Training was broken into two phases: pretraining using convolutional denoising autoencoders (CDAE) on over one million simulated physiological signals and fine-tuning using transfer learning on a collected set of real-world data, Fig. 1
The proposed algorithm performs collaborative multitask feature learning for two correlated tasks, input signal quality assessment and event detection (AF presence)
Summary
Wearable devices are increasingly used in cardiology for out-ofthe-clinic healthcare monitoring[1]. Poh et al.[8] proposed a method for dense CNN to distinguish between noise, sinus rhythm, ectopic rhythms, and AF across an ensemble of three simultaneously collected PPG signals These AF classification methodologies do not consider joint estimations of signal quality assessment or explore transfer learning to boost discriminatory power. Exploring transfer learning for AF detection appeals to biomedical research given the common challenge of limited access to large labeled cohorts. To address this gap, we present DeepBeat, a method for the detection of AF from wrist-based PPG sensing. Our method combats the unique noise artifact problem common in AF
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
