Group-Personalized Regression Models for Predicting Mental Health Scores From Objective Mobile Phone Data Streams: Observational Study

Abstract

BackgroundObjective behavioral markers of mental illness, often recorded through smartphones or wearable devices, have the potential to transform how mental health services are delivered and to help users monitor their own health. Linking objective markers to illness is commonly performed using population-level models, which assume that everyone is the same. The reality is that there are large levels of natural interindividual variability, both in terms of response to illness and in usual behavioral patterns, as well as intraindividual variability that these models do not consider.ObjectiveThe objective of this study was to demonstrate the utility of splitting the population into subsets of individuals that exhibit similar relationships between their objective markers and their mental states. Using these subsets, “group-personalized” models can be built for individuals based on other individuals to whom they are most similar.MethodsWe collected geolocation data from 59 participants who were part of the Automated Monitoring of Symptom Severity study at the University of Oxford. This was an observational data collection study. Participants were diagnosed with bipolar disorder (n=20); borderline personality disorder (n=17); or were healthy controls (n=22). Geolocation data were collected using a custom Android app installed on participants’ smartphones, and participants weekly reported their symptoms of depression using the 16-item quick inventory of depressive symptomatology questionnaire. Population-level models were built to estimate levels of depression using features derived from the geolocation data recorded from participants, and it was hypothesized that results could be improved by splitting individuals into subgroups with similar relationships between their behavioral features and depressive symptoms. We developed a new model using a Dirichlet process prior for splitting individuals into groups, with a Bayesian Lasso model in each group to link behavioral features with mental illness. The result is a model for each individual that incorporates information from other similar individuals to augment the limited training data available.ResultsThe new group-personalized regression model showed a significant improvement over population-level models in predicting mental health severity (P<.001). Analysis of subgroups showed that different groups were characterized by different features derived from raw geolocation data.ConclusionsThis study demonstrates the importance of handling interindividual variability when developing models of mental illness. Population-level models do not capture nuances in how different individuals respond to illness, and the group-personalized model demonstrates a potential way to overcome these limitations when estimating mental state from objective behavioral features.

Sensing Apps and Public Data Sets for Digital Phenotyping of Mental Health: Systematic Review

“…Positioning and inertial SituMan [39] Correlation analysis and mental state classification Semantic locations, physical activity, sociability Positioning, inertial, virtual, and ambient EmotionSense [40] Correlation analysis Sociability, mobility, physical activity, device usage Positioning, inertial, virtual, and ambient StudentLife [41] Correlation analysis Physical activity, mobility, and sociability Positioning, inertial, and ambient Undefined [42] Mental state prediction Mobility Positioning AMoSS [43] Mental state classification Mobility Positioning and virtual eB2 [44] Raw data collection It does not infer information Positioning, inertial, virtual, and ambient EARS [45] Mental state classification Posture/position of body when sleeping Inertial and ambient SleepGuard [46] Mental state classification It does not infer information Virtual Moment [47] Mental state classification It does not infer information Virtual TypeOfMood [48] Raw data collection It does not infer information Positioning, inertial, virtual, and ambient RADAR-base [49] Correlation analysis and mental state classification Physical activity, mobility, sleep, sociability Positioning, inertial, and ambient SHADO [50] Raw data collection It does not infer information Positioning, inertial, virtual, and ambient InSTIL [51] Correlation analysis Physical activity Positioning Lamp [52] Correlation analysis Mobility, sociability, context of daily life (eg, duration of sleep) Positioning, inertial, virtual and ambient SOLVD [53] Mental state classification Physical activity, mood, sociability, sleep Inertial, virtual, and ambient STDD [54] Mental state classification Sociability and mobility Positioning, virtual, and ambient Moodable [55] Mental state classification Mood, stress level, and well-being…”

Dynamic clustering via branched deep learning enhances personalization of stress prediction from mobile sensor data

“…However, the way people experience mental conditions and specifically stress levels vary significantly from individual to individual 27 . To address this challenge, personalized models have been introduced and explored 17 , 28 – 30 . Techniques involving personalized parameters have been used for a wide range of applications, from intelligent sensing and health status monitoring 31 , to next-word prediction on mobile devices 32 .…”

Toward the Personalization of Biceps Fatigue Detection Model for Gym Activity: An Approach to Utilize Wearables’ Data from the Crowd

“…The rationale behind the personalization is to prioritize the crowd's data from whom the test subject is similar. Therefore, we study the physical traits of the crowd to improve personalization of the cross-subject model by adding weights to the training data entries based on the intersubject similarities between the test subject and the subjects from the crowd [28][29][30][31][32]. We believe that combining similarities measure and personalization can form a personalized model which has a better performance than the cross-subject model and consumes the test subject's data less than the subject-specific models.…”