It seems like everyone is wearing a fitness device these days. The Apple watch, Fitbit, Oura ring and many others aim to monitor our bodies to keep us healthy. Most of these devices measure heart rate and step count while, increasingly, some go further to try to detect disease. Our paper explores how we can combine wearable devices with deep learning “artificial intelligence” methods for the real time detection of a common abnormal heart rhythm known as atrial fibrillation.
Atrial fibrillation affects upwards of one in five people in their lifetime1. It describes the chaotic electrical activity of the top chambers of the heart, the atria. It can cause symptoms of palpitations but, if it goes unnoticed as it commonly does, it can be especially dangerous through its risk of stroke2. Stroke is a condition where the brain is damaged by a bleed or a blood clot which causes weakness down one side of the body and potentially makes it difficult to swallow and speak. Stroke is one of the largest sources of healthcare expenditure as patients often require intensive rehabilitation for months or years afterwards3. Detecting atrial fibrillation early and preventing stroke could save millions of dollars and prevent major disparities in care4.
Wearable devices are ideally suited to detect atrial fibrillation because they can measure heart rate by intermittently shining light into the skin. A sensor in the wearable device detects returning light of different intensities as the blood from each beat passes beneath it. A major unsolved problem in the field is the challenge of noise, created by the movement of the wrist during everyday activities.
In some areas of research, like biomedical applications, it is very difficult to collect large, well-documented datasets due to the cost of data collection, data labelling, and patient privacy. This can limit development and accessibility. Thus, the ability for AI algorithms to learn effectively from smaller datasets or unlabeled datasets is critical.
In this paper, we use a technique called transfer learning that helps solve the problem of small training data sets. The goal is to build models that perform well, even when the data sources for training are different. With the expansion of deep learning to medical applications, transfer learning has become integral to its success.
In this work, we aggregate over one million simulated unlabeled physiological signals and curated datasets of over five hundred thousand labeled signals from over one hundred individuals from three different wearable devices to train and test our approach, DeepBeat.
We found the following:
 Using a multi-task learner for two correlated tasks, rhythm estimation and signal quality estimation, allows for joint learning of both tasks and improved performance.
 The use of convolutional denoising autoencoders (CDAE) for unsupervised learning as a pre-training technique is beneficial and leads to better distinction in learned class representations.
In a prospective validation study, DeepBeat maintained high detection rates (sensitivity 0.98, specificity 0.99, and F1-score 0.93). Our deep learning approach serves as a foundational step towards early detection of atrial fibrillation and a reduction in the financial and emotional trauma of stroke.
- Lloyd-Jones Donald M. et al. Lifetime Risk for Development of Atrial Fibrillation. Circulation 110, 1042–1046 (2004).
- Wolf, P. A., Abbott, R. D. & Kannel, W. B. Atrial fibrillation as an independent risk factor for stroke: the Framingham Study. Stroke 22, 983–988 (1991).
- Buntin, M. B., Colla, C. H., Deb, P., Sood, N. & Escarce, J. J. Medicare spending and outcomes after postacute care for stroke and hip fracture. Med. Care 48, 776–784 (2010).
- Reynolds, M. R. & Essebag, V. Economic burden of atrial fibrillation: Implications for intervention. Am. J. Pharm. Benefits 58–65 (2012).