UCLA

Depression and anxiety, the two most common mental health conditions, consist of heterogeneous combinations of altered physiology, behaviors, and emotional states. The lack of scalable, objective methods to measure these dimensions is a major obstacle to elucidating underlying mechanisms and improving care for those who experience these conditions. Here, I show that such measurements can be obtained from sensor data passively and continuously collected from smartphones and smartwatches. First, I demonstrate the feasibility of collecting such data from a large sample of diverse individuals who completed an extensive array of self-report measures while providing objective digital phenotyping data. Second, I discuss various strategies for processing, cleaning, and analyzing these data, all of which I implemented in a robust processing pipeline. Finally, I identify associations between these digital phenotyping features and self-reported measures of depression and anxiety. From 12-months of sensor data provided by 3,751 consenting participants (diverse in age, sex at birth, race, ethnicity, and severity of depression symptoms), I extracted 427 digital features, yielding assessments of physiology, behavior, and emotional expression; evaluations in the latter category included novel measurements of facial expression, typing sentiment, and speech. Most digital features were significantly associated with overall levels of depression and anxiety. I identified stronger associations, however, by analyzing the features in relation to 19 specific symptom dimensions (constructs), which were aggregated from 69 individual symptoms through consensus of clinical experts. While most features were associated with multiple symptom constructs, each construct displayed distinctive patterns of association to different features. Notably, suicidality, perhaps the most consequential dimension of depression, was especially strongly associated with features that scalably and longitudinally quantify variability in facial expression, typing sentiments, heart rate, and sleep. Additionally, I identified subsets of feature-construct associations that differ substantially across groups defined by age, sex, or symptom severity. My results demonstrate that consumer devices can provide an objective and scalable framework for phenotyping depression and anxiety and dissecting the complexity of these conditions, key steps toward precision mental health.