On Detecting COVID-Risky Behavior from Smartphones

Thomas Hartvigsen, Walter Gerych, Marzyeh Ghassemi

Keywords: digital health, time series, human activity recognition

TL;DR: We show that some COVID-risky activities are somewhat predictable from smartphones. Detecting such behavior could shorten the feedback loop between public health messaging and population response. Privacy concerns temper these benefits.

Abstract: Detecting risky behavior using smartphones and other mobile devices may help mitigate the spread of infectious diseases. However, the privacy concerns introduced by individualized activity recognition may counteract potential benefits. As an example, consider a public health official gauging their message's efficacy. Machine learning and data mining methods may help them understand how their local population's behavior changes, but aggressive surveillance can severely hinder individual privacy and cause disproportionate harm to disadvantaged groups. In this work, we benchmark a series of machine learning algorithms predicting high-risk behaviors---going to bars and gyms, attending parties, and riding on buses---given only low-level smartphone sensor data. We find that models trained to perform these challenging tasks are largely unreliable and should be avoided in practice, though their predictions are significantly better than random.