An interdisciplinary team of researchers at the University of Michigan is exploring how to predict human resilience, providing insights into anything from athletic and academic performance to disease susceptibility and injury recovery. The effort, called “Enhancing Mechanisms of Human Resilience for Student Success and Well-Being,” has been provided with seed funding from the U-M Biosciences Initiative. Morris Wellman Assistant Professor in CSE Mosharaf Chowdhury, one of five Co-Principal Investigators on the project, will explore new methods to keep user data private during the study using a technique called federated machine learning.
Modern technology has brought us closer than ever to understanding and predicting human performance in a variety of contexts. The U-M researchers are interested in determining the various factors that impact a person’s performance on specific tasks, like exams or job interviews, and how well they withstand unforeseen circumstances, such as an illness.
In the case of something like an exam, Chowdhury explains, a student’s sleeping patterns and time spent studying could be used to predict their score. In the case of physical wellness, factors including their nutrition and fitness could be used to predict their ability to recover from injuries. All of these things together constitute what the team considers “human resilience.”
“In this study, we will collect a range of physiological data and combine it with survey data,” Chowdhury says. “We’ll use it to understand the overall health of the study population and to predict how they will perform in different scenarios.” The data being used for the initial phase of the study includes Fitbit data, such as heart rate and activity levels, blood tests, and recurring surveys collected via an app from 2500 volunteer undergraduates.
The app-collected data will be the focus of Chowdhury’s contributions to the project. Currently, studies and products that make use of user data from personal devices rely on centralizing the data on a server in order to run necessary analytics. This offloading of user data brings privacy concerns, as any security risks faced by the server become shared risks for all of the users.
The human resilience team will build two data analytics frameworks – one for reference with a typical centralized design, and one designed by Chowdhury to leverage an emerging technique called federated machine learning. The latter performs model training on the participant data without moving the data itself off of user devices. The training becomes distributed across all of the user devices, and only the resulting insights from the data are then shared with the research team.
“Under this model, there is no requirement to collect data in any central place; it will remain distributed,” he explains. “You will still ideally be able to perform most or all machine learning tasks.”
This setup allows the researchers to compare the two methods directly, identifying any limitations arising from the young federated learning technology.
“This will give us a proof of concept showing that, even in such a complex problem, where all of the answers to user questions aren’t known ahead of time, we can allow people to respond to these private queries without ever giving up ownership of their data,” says Chowdhury. “At the end of the day, researchers and companies just need the insights from data, and not the data inself.”
The Co-PIs of the project are Julia Lee Cunningham (Ross School of Business), Muneesh Tewari (Michigan Medicine), Mosharaf Chowdhury (College of Engineering), Kira Birditt (Institute for Social Research), and Sung Won Choi (Michigan Medicine). Additional co-investigators include Walter Dempsey (School of Public Health), Stephen M. Cain (College of Engineering), Joel Gagnier (Michigan Medicine), Natalie Colabianchi (School of Kinesiology), and Ron Zernicke (Michigan Medicine). The team spun out from a U-M Biosciences Initiative Ideas Lab workshop in 2019. After a delay due to COVID-19, the project launched in early 2021.