top of page

ML for COVID-19: Can AI give you an alert indicating a viral infection before you feel symptoms?

Imagine you could look at your phone and learn, based on data that it and your smart watch collected, that it forecasts your health score dropping more than 40 points (on a 100 point scale) tomorrow, with a 95% confidence interval of +- 13 points. The forecast is validated on prior data, and comes with suggestions of things you might do tonight to mitigate that drop. Instead of thinking “I’m probably just tired and stressed” you might consider a change: Perhaps you decide not to push to finish work you had planned tonight; instead, you go to bed early, hoping to boost your immune system’s natural abilities to keep you well by getting extra sleep. You might also cancel gathering with people – to reduce risk of exposing them, even though you have no fever or cough/sneezing. This early action might help reduce the severity of or prevent your illness, or even, with such advanced notice, help prevent others from exposure.

The underlying technology required to implement an AI-based health/sickness forecast was first demonstrated to have the accuracy described above in the MIT Media Lab’s Affective Computing Research group in a study using 1,895 days of data collected with sensors and smartphones worn by 69 college students. The approach achieved less than 13% mean absolute error in estimating tomorrow’s health score ranging on a scale from 0=sick to 100=healthy[1].

After further studies and review, the U.S. Government’s BARDA Division of Research, Innovation and Ventures (DRIVe) funded Empatica, an MIT spin-out co-founded by J-Clinic investigator Rosalind Picard, to develop a new tool that will alert wearers about a serious respiratory infection before any symptoms appear. Empatica provides an FDA-cleared smartwatch called “Embrace” that collects continuous physical, autonomic and skin temperature data, while running on-board real-time AI algorithms and sending alerts through a paired smartphone to the user and to designated caregivers. Empatica’s watch and AI-based analytics have been previously cleared by FDA for use in neurology for detecting the most life-threatening type of seizure.

Working with Dr. Jeffrey Shaman’s team at Columbia University, Picard and Empatica are now developing machine learning to identify the first inflammatory response biomarkers to a virus using the wearable sensor data. The Embrace watch is unique in continuously providing physiological patterns including autonomic changes in electrodermal activity caused by the sympathetic nervous system (SNS) without parasympathetic influence. Central sympathetic activity has been established to have a direct impact on inflammatory cytokines and lymphoid tissue is highly innervated by sympathetic nerve fibers, with sympathetic nerve terminals located close to immune cells. Patterns in the watch data are analyzed continuously for estimating changes related to the earliest phases of infection.

While the BARDA-funded work focuses on biomarkers for influenza infections, the team has now begun focused research on SARS-CoV-2 viral infections. The current challenge is to collect data and harness the power of AI-machine learning together with continuous wrist-worn physiological data to help detect and alert to the very earliest signs of infection while a person is asymptomatic. The objective data can provide information related to stress, slow-wave sleep, temperature, activity, and autonomic function that may shed insight into why, with the same virus load, different people have vastly different responses to the virus.

Working together with partners on the front lines, we hope to bring the power of AI to enable better health outcomes in defense against COVID-19.


[1] The first paper describing the machine-learning implementation and test details (also comparing this approach to several other machine learning models) was published in Jaques et al. 2017: "Predicting Tomorrow's Mood, Health, and Stress Level using Personalized Multitask Learning and Domain Adaptation," Proc. of Machine Learning Research, 48, 17-33. August 2017.

bottom of page