Passive COVID-19 Assessment using Machine Learning on Physiological and Activity Data from Low End Wearables

Abstract

COVID-19 has now infected over 165 million people and killed over 3.5 million people. While public health interventions have reduced its spread and vaccines are being deployed, passive detection methods are needed to detect infections and early track its resurgence. Wearables that are widely owned can gather various physiological and activity data, presenting an opportunity to detect COVID-19 unobtrusively. COVID-19 infection causes deviations in the vital physiological signs and activity patterns of infected users. However, similar deviations of these same variables can also be affected by non-COVID factors, confounding the signals. In this paper, we investigate the feasibility of predicting COVID-19 infection to detect abnormalities in heart rate, activity (steps), and sleep data available on low-end wearables by using machine learning. Prior work utilized data such as oxygen saturation that is only available on clinical-grade equipment or expensive wearables. We extracted 43 statistical features (standard deviation, mean, slope) and behavioral (min/max/avg length of sedentary and active bouts, sleep duration, no. of awake/asleep/restless samples) from wearable sensor data. We classified these features using machine learning classification and anomaly detection algorithms. Physical activity features were the most predictive (min length of the sedentary and active bout), yielding an AUC-ROC of 78% [specificity=74%, sensitivity=69%] when classified using Gradient Boosting Machines (GBMs). We also found that sleep irregularities had low discriminative performance. COVID-19 detection using inexpensive wearables can facilitate population-level interventions.