Using Digital Phenotyping for Depression Screening in Community-Dwelling Older Adults: Bayesian Multilevel Hurdle...

With the rapidly aging population, mental health among older adults has received growing attention. Although the likelihood of experiencing depressive symptoms is higher in late adulthood, older adults are more reluctant to visit a clinic due to the stigma surrounding mental health issues, and many remain undiagnosed and untreated. Digital phenotyping has emerged as a promising approach to mitigate this problem. Longitudinal monitoring via wearable devices can facilitate the timely identification of depressive symptoms in older adults. However, there has not been sufficient investigation to develop a machine learning approach that accounts for between-person and within-person characteristics.

This study aimed to investigate the utility of active and passive digital phenotyping data collected via wearable devices for monitoring the probability and severity of depressive symptoms. Specifically, we applied multilevel hurdle modeling within a machine learning framework to enable efficient depression screening in the general population, with a focus on community-dwelling older adults.

We analyzed 1011 cases reported by 147 older Korean adults for 2 years. Participants were asked to complete the 9-item Patient Health Questionnaire (PHQ-9) items in our mobile app during the last week of each month. In addition to the annual in-person data collection, we also collected active and passive sensing data from participants via smartphones and smartwatches. For dimensionality reduction on 44 features, parallel analysis and principal component analysis were used. With the extracted 6 principal components (PCs), a Bayesian multilevel hurdle model was used.

When constructing PCs, the weekly stress rating from active data and sleep-related features from passive data were the top 5 contributing features. Among the 6 PCs, the PC consisting of low psychological distress and high social support was significantly associated with depressive symptoms in community-dwelling older adults. This Bayesian multilevel hurdle model showed good performance in screening for depressive symptom severity (R2=0.53) and in distinguishing between those with and without symptoms (area under the receiver operating characteristic curve=0.88 and F1-score=0.75) on the test data. The between-person variance was larger than the within-person variance, especially in explaining the probability of depressive symptoms.

In mental health screening, active and passive digital phenotyping data can be used in conjunction with traditional clinical screening tools to monitor depressive symptoms among community-dwelling older adults. Dimensionality reduction via parallel analysis and principal component analysis can help identify latent risk profiles. Given the nested data structure and heterogeneity in depressive symptoms, a Bayesian multilevel hurdle model within a machine learning framework may be helpful for depression screening. Overall, digital phenotyping can be a useful tool for personalized, within-person health tracking, even after accounting for substantial between-person variance. We recommend future work to address data imbalance to further strengthen this approach.