Personalized prediction of negative affect in individuals with serious mental illness followed using long-term multimodal mobile phenotyping.

Abstract

Heightened negative affect is a core feature of serious mental illness. Over 90% of American adults own a smartphone, equipped with an array of sensors which can continuously and unobtrusively measure behaviors (e.g., activity levels, location, and phone usage patterns) which may predict increases in negative affect in real-time in individuals' daily lives. Sixty-eight adults with a primary mood or psychotic disorder completed daily emotion surveys for over a year, on average (mean 465 days; total surveys = 12,959). At the same time, semi-continuous collection of smartphone accelerometer, GPS location, and screen usage data, along with accelerometer tracking from a wrist-worn wearable device, was conducted for the duration of the study. A range of statistical approaches, including a novel personalized ensemble machine learning algorithm, were compared in their ability to predict states of heightened negative affect. A personalized ensemble machine learning algorithm outperformed other statistical approaches, achieving an area under the receiver operating characteristic curve (AUC) of 0.72 (for irritability) -0.79 (for loneliness) in predicting different negative emotions. Smartphone location (GPS) variables were the most predictive features overall. Critically, there was substantial heterogeneity between individuals in the association between smartphone features and negative emotional states, which highlights the need for a personalized modeling approach. Findings support the use of smartphones coupled with machine learning to detect states of heightened negative emotions. The ability to predict these states in real-time could inform the development and timely delivery of emotionally beneficial smartphone-delivered interventions which could be automatically triggered via a predictive algorithm.