AI predicts college student stress from phone sensor and questionnaire data

College students lead stressful lives. They’ve got assignments to complete and extracurriculars to attend, not to mention tests to prepare for and job applications to submit. Unfortunately for them, the negative health effects of stress are well-documented. If left untreated, it can cause cardiovascular diseases, affect memory and cognition, and even suppress the immune system.

To help suss out the outsized contributors to social and academic pressure, researchers at the College of Computer Science at the University of Massachusetts turned to AI, which they used to predict stress levels — below median, median, or above median — from questionnaire and smartphone sensor data. They report that their model achieved state-of-the-art performance, obtaining a 45.1% improvement compared with the baseline on a data set of student sleep patterns, activity, conversation, location, information regarding mental health (like stress levels), and more.

They describe their work in a paper (“Personalized Student Stress Prediction with Deep Multitask Network“) published on the preprint server Arxiv.org.

“With the growing popularity of wearable devices, the ability to utilize physiological data collected from these devices to predict the wearer’s mental state, such as mood and stress, suggests great clinical applications, yet such a task is extremely challenging,” wrote the coauthors. “With the induction of high-quality robust sensors in wearables like Fitbit, Apple Watch, and smartphones, efficient collection of physiological and behavioral data with reasonable accuracy has become affordable.”

The researchers’ AI system — Cross-personal Activity LSTM Multitask Auto-encoder Network, or CALM-Net — considers data as time-series (i.e., taken at successive equally-spaced points in time) and can identify temporal patterns contained within it. Additionally, it offers the ability to personalize models and incorporate time-series information, which improves performance as the number of students increases. And it infers and measures things like day of the week, sleep rating, sleep duration, and time to next assignment deadline.

The corpus on which the model trained — StudentLife — was conducted at Dartmouth and facilitated by an Android app. Each day over a 10-week period, it recorded stress data from 48 students in the form of ecological momentary assessments (i.e., real-time responses to questionnaires), which it paired with passive sensing data like activity (walking, running), audio (silence, voice, noise), phone charge level, and phone lock status as often as every 10 seconds.

In tests, CALM-Net improved upon of the performance of the two state-of-the-art models to which it was compared. The researchers theorize this was a direct result of its ability to consider data as time-series. “The ability of CALM-Net to incorporate granular temporal information and high-level covariates, along with an architecture which is capable of deciphering personalized patterns for each student without overfitting, contributes to its high performance,” they wrote. “[This approach] improves the performance of all evaluated models, showing that stress indicators can generally be better modeled using personalized layers.”