.png)
Symposia
Technology/Digital Health
Geneva K. Jonathan, Ph.D. (she/her/hers)
Massachusetts General Hospital
Boston, MA, United States
Vincent Holstein, MD (he/him/his)
Resident
Massachusetts General Hospital
Boston, MA, United States
Michael Armey, PhD
Associate Professor of Research
Warren Alpert Medical School of Brown University
providence, RI, United States
Jukka-Pekka Onnela, DSc
Associate Professor of Biostatistics; Co-Director of the Master of Science in Health Data Science Program
Harvard T.H. Chan School of Public Health
Boston, MA, United States
Hilary Weingarden, Ph.D. (she/her/hers)
clinical Research Director
HabitAware
Arlington, MA, United States
Background: Body Dysmorphic Disorder (BDD) is a debilitating psychiatric condition characterized by distressing preoccupations with perceived appearance flaws, leading to social withdrawal, functional impairment, and increased risk of suicidal ideation (SI). Traditional approaches for monitoring clinical deterioration in BDD rely on self-reports and periodic clinician-administered assessments, which can miss acute changes in risk due to infrequent administration and recall biases. This study tests the feasibility of using smartphone sensor data and self-reported demographic variables to predict daily symptom severity and suicide risk.
Methods: A sample of 82 participants with clinically significant BDD symptoms completed ecological momentary assessments (EMA) over 28 days, reporting levels of suicidal ideation, avoidance behaviors, and time spent on BDD-related concerns. Concurrently, passive smartphone sensor data were collected using the Beiwe app. Machine learning models, including Random Forest and Elastic Net regression, were trained to predict same-day symptom severity using GPS, accelerometer, and demographic data. Model performance was evaluated using mean absolute error, Pearson correlations, and permutation testing.
Results: Random forest models using time- and random-split validation outperformed baseline models across all outcomes (max SI, mean SI, max avoidance, mean avoidance, time-spent on BDD-related behaviors). Step count and demographic factors, such as education and living situation, were as strong predictors of symptom severity. Models trained on one set of users (user-split scenario) did not generalize well to new users.
Conclusions: This study provides initial evidence that passive smartphone sensor data and demographic information can be leveraged to detect same-day clinical worsening in BDD. Findings show the potential for integrating digital phenotyping into just-in-time adaptive interventions, offering a scalable, low-burden approach to monitoring symptom changes in real-world settings. Future research should explore the integration of these predictive models into digital mental health interventions and clinical decision-making.