Predicting dimensions of depression from smartphone data

Depressive disorders are highly prevalent but demand nuanced personalized treatment that traditional approaches in psychiatry cannot address. This gap has prompted a surge of interest in leveraging digital technology, particularly smartphones, for remote monitoring to enhance outpatient care. This study utilizes the BRIGHTEN dataset to construct interpretable prediction models for overall depression severity, measured by PHQ-9, and various depression dimensions using a factor modelling approach. Our factor model unveils a three-factor solution encompassing mood, somatic, and concentration/psychomotor-related factors. Machine learning models effectively predict both the PHQ-9 scores and individual factors, with feature importance methods analyses underscoring the influence of the PHQ-2 scale and communication-related features. These findings are corroborated by models trained on data subsets. Through nested multi-level models, we identify between-subject effects for the PHQ-2 and select communication-related features, along with within-subject effects for these features. In summary, this study underscores the robust predictive capacity of ecological momentary assessments and highlights features of potential relevance for future investigations, such as communication-related features. We advocate for future studies to assess the cost-effectiveness and intervention potential of these models.