Detecting dynamic fluctuations in emotion, mood, and functioning: A digital phenotyping approach to clinical monitoring in bipolar disorder

Project Summary/Abstract Digital phenotyping, the in-situ quantification of individual-level phenotypes using data from digital devices, is a promising new tool for measuring patterns of emotion and mood in the context of individuals lived experience. Our work using ecological momentary assessment (EMA), an interactive form of digital phenotyping, shows

that patterns in emotional valence and arousal associate with and predict mood symptom severity. Specifically, mean levels, greater day-to-day fluctuations (instability), and large increases or decreases (anomalies) in emotion are associated with heightened risk. However, EMA is burdensome if administered frequently,

highlighting the need for complimentary passive digital phenotyping methods that unobtrusively measure emotion patterns. Our team has developed an app and a series of machine learning algorithms, PRIORI, that samples the ambient audio every 15 minutes and produces computational estimates of emotional valence and

arousal that correlate with EMA measures. The goal of the present study is to investigate how to mitigate the burdensome participation requirements of EMA, and substantially improve the ability to predict mood symptoms using passive (PRIORI) technology. Individuals (n=160) with an established pattern of mood

instability from a prospectively studied cohort of bipolar individuals will enroll into a six-month digital phenotyping protocol including continuous passive monitoring using PRIORI. In a measurement-burst protocol one week per month, participants will complete EMA self-reported mood (2xday), self-reported emotion

valence and activation (5xday), and a clinician interview to assess manic and depressive symptoms. Outside of these bursts, participants will complete weekly EMA ratings of mood. Our central hypothesis is that patterns of emotional valence and arousal assessed via digital phenotyping will predict mood severity in BD. We will test

whether the prediction of mood severity using EMA data will be enhanced by including measures derived from PRIORI. Given that EMA data may not always be present (while PRIORI data are), we will also test whether including emotion estimates from PRIORI during weeks when EMA is not present improves later mood

prediction. Our specific aims test independent hypotheses focused on: (1) mean levels of emotional valence and arousal, (2) day-to-day fluctuations (instability) in emotional valence and arousal, and (3) large deviations in emotional valence and arousal from one’s own average (anomalies). For risk mitigation and alternate

strategies, we will examine the extent to which contextual variables (noise from the environment, location, social factors) and psychosocial functioning influence predictive models. This project will support the development of fine-grained and quantitative behavioral assessment tools to evaluate dysfunction in the

trajectories of mental illness (NIMH Obj 2.2) and elucidate additive, interactive combinations of data that can identify dynamic risk (NIMH Obj 2.2A). Further, study findings have the potential to provide critical data on personalized models of psychopathology with a focus on the emotion and mood dynamics.