Multivariate Time Series Subgroup Analysis Using Smartphone Passive Sensing for Mental Health Assessment Minsik Lee Ulsan National Institute of Science and Technology

Abstract

Traditional mental-health assessments rely heavily on surveys, which are collected infrequently, im- pose user burden, and often fail to capture rapid changes in everyday psychological states. Passive smart- phone sensing offers continuous and low-effort observation of real-world behavior, but it remains unclear whether behavioral structure alone—without clinical labels—can reveal meaningful mental-health pat- terns. This study addresses this challenge by constructing a rhythm-sensitive weekly representation of passive sensing signals and identifying latent behavioral states through unsupervised clustering. Using 2,659 weekly samples from 536 participants, five distinct behavioral states were discovered, each characterized by differences in circadian regularity, sleep–wake stability, daytime engagement, and the timing of nighttime smartphone use. States with weakened rhythmic structure or concentrated nocturnal activity showed markedly elevated symptoms of depression, anxiety, insomnia, and stress. Feature-level analyses consistently highlighted 24-hour rhythm amplitude, sleep-timing variability, day- time–nighttime contrast, and weekday–weekend structure as the most discriminative behavioral markers. These domains also emerged as dominant predictors in multivariate importance models, reinforcing their central role in weekly behavioral organization. Temporal transition analysis further revealed two distinct vulnerability mechanisms: a short-lived but severely irregular rhythm state, and an acute disturbance state marked by brief spikes in nighttime and weekend engagement. In contrast, stable states with strong rhythmicity showed low symptom burden and high week-to-week persistence. Taken together, the findings demonstrate that passive sensing alone can recover meaningful behav- ioral states that align closely with psychological risk. The rhythm-aware representation developed in this study provides a label-efficient framework for continuous monitoring, early detection of behav- ioral disruption, and scalable real-world applications such as adaptive mental-health interventions and state-based just-in-time support systems.