Using data processing to understand inconsistency in smartphone behavior among patients with serious mental illness: Results of a digital phenotyping biomarker study

Abstract

Background

Smartphone-based digital phenotyping can provide novel transdiagnostic markers of mental illness including circadian routines and anhedonia. In proposing transdiagnostic digital phenotypes for circadian routines and anhedonia in depression and bipolar disorder patients, this paper explores their derivation, comparison to naive models, and replicability across two different research sites/teams.

Methods

84 participants (bipolar disorder, depression, controls) used the mindLAMP app for 12 weeks to capture digital phenotypes on their personal smartphones. mindLAMP was used to deliver surveys about mood symptoms while collecting device acceleration, geolocation, and screen on/off state. Participant chronotype was derived from this sensor data. Within-participant and between-participant models were created to assess how time-varying features collected through digital phenotyping could predict weekly anhedonia survey responses.

Results

Within-person models outperformed between-person models in predicting anhedonia. Chronotype was the strongest predictor of weekly anhedonia scores as indicated by Shapley scores. Shapley scores also revealed that many of the time-varying predictor variables are significant but differ in their direction of action.

Discussion

This analysis reveals the meaningful but potentially misleading nature of digital phenotyping signals. Results suggest that each participant has a unique set of relationships between time-varying digital phenotype variables; therefore, it is challenging to predict trends between participants. Bayesian models, with appropriate population priors, may offer the next step for improving the potential of personalized digital phenotyping insights.