Within-person prediction of depressive symptom change using year-long Screenome data and CES-D assessments

A new study from Stanford University, published on arXiv, demonstrates that AI can predict whether an individual's depressive symptoms will worsen, stay stable, or improve over the next two weeks with remarkable accuracy. The research, led by Merve Cerit and colleagues, leverages the Screenomics platform—a tool that captures phone screenshots every five seconds—to build a digital phenotype of user behavior. Over a year, the team collected over 100 million screenshots from 96 adults, pairing this data with fortnightly CES-D (Center for Epidemiologic Studies Depression Scale) assessments for a total of 2,002 observations.

The study framed prediction as a within-person classification task, using XGBoost to achieve an AUC of 0.906 for detecting crossings of established CES-D severity bands, and 0.755 for changes relative to each participant's own variability. The model generalized to unseen individuals with an AUC of 0.821. Crucially, the most significant predictor beyond the latest CES-D score was each person's typical symptom level—without it, worsening transitions went undetected. Screenome-derived features revealed prodromal patterns like escalating social media use, fragmented device engagement, and changes in overnight activity, though with substantial individual heterogeneity. This work establishes a proof-of-concept for monitoring systems that could identify individuals approaching clinical deterioration before symptoms reach a crisis point, enabling earlier and more targeted mental health care.