JMIR Publications

ABSTRACT

Background:

Following the rapid population aging, mental health for older adults has received growing attention. The likelihood of experiencing depressive symptoms becomes higher in late adulthood. However, older adults are reluctant to visit a clinic with their mental health issues. As a result, many remain undiagnosed and untreated. Digital phenotyping can help solve this problem. Although longitudinal monitoring using wearable devices can be helpful for promptly identifying the risk of depression in older adults, there has not been sufficient investigation to develop a machine learning approach that differentiates between-person and within-person changes.

Objective:

This study aims to investigate whether and how collecting active and passive digital phenotyping data through wearable devices can help predict the risk of depression in older adults. In particular, this study considers multilevel modeling as part of a machine learning approach to monitor the risk of depression in older adults, distinguishing between-person and within-person changes.

Methods:

Using Bayesian models, we analyzed 1,011 cases reported by 147 older Korean adults. Participants were asked to complete the PHQ-9 items in our mobile app during the last week of each month. Participants also visited annually to participate in our in-person data collection on their socio-demographic, physical, and psychological characteristics. Beyond the annual in-person data collection, we collected active data, such as daily mood and weekly stress, through our mobile application. Passive-sensing data, including step counts and sleep logs, were collected via a smartwatch. Bayesian multilevel models with continuous and binary outcomes, respectively, were employed to predict depression in older adults.

Results:

Bayesian multilevel machine learning models yield good model performance. Adding active data and passive sensing data demonstrates high AUCs and R-squared values in Bayesian models. There were significant differences in depressive symptoms at the between-person level and the within-person level. The mean and standard deviation of daily mood, the mean and standard deviation of weekly stress, and the average minutes of deep sleep were significantly associated with older adults' depressive symptoms in the month. Participants' age, anxiety, and self-report regular sleep hours were also significantly associated with depressive symptom severity.

Conclusions:

In mental health screening, collecting active and passive digital phenotyping data can be used in conjunction with traditional clinical screening tools to identify high-risk groups for depression and to monitor its severity in the general aged population.