Transfer Learning for the Prediction of Atrial Fibrillation and Sleep Apnea
Inexpensive and highly portable medical sensing technology has the potential to deliver savings for healthcare providers, through screening and early detection. However, the data necessary to support machine learning models on these newly developed sensor modalities - sensor streams together with time-aligned, expert-annotated labels - is often difficult and expensive to obtain. By contrast, high-quality data suitable for machine learning on better-established modalities, such as electrocardiogram (ECG), is abundant and often free. We propose a method allowing for the use of such easy-to-come-by data in building models on photoplethysmography (PPG) sensor modalities. With this method, data requirements are reduced to streams of PPG data time-aligned with readings from ECG together with labelled outcome for ECG. We demonstrate the method by building models for atrial fibrillation and sleep apnea based on data from a wrist-worn PPG sensor, with the only labels coming from publicly available ECG data. We find that the models developed using the transfer learning approach outperform models trained directly on the PPG sensor and are competitive with state-of-the-art ECG-based models.