Methods for characterizing circadian physiology from wearables

Abstract

Non-invasive data collection in real-world settings with wearables provides a new opportunity for characterizing daily physiology. However, accurate and efficient characterization remains an open problem because the complex autoregressive noise of the data makes it challenging to use a simple and efficient method for inference of clock proxies, least squares method. In this talk, we will introduce a simple approximation that alters the noise structure and thus enables one to use the least squares method. We will show its usefulness for real-time personalized fever detection in cancer patients.