MODELING CONTINUOUS-TIME NETWORK DYNAMICS OF PASSIVE SENSING AND ECOLOGICAL MOMENTARY ASSESSMENT DATA TO CHARACTERIZE TREATMENT RESPONSE IN DIGITAL MENTAL HEALTH CARE

Rapid advances in patient-generated health data platforms have expanded opportunities for digital psychiatric care through passive smartphone sensing and ecological momentary assessment (EMA) data. Yet, prior studies have largely focused on bivariate associations rather than modeling the full multivariate network. Furthermore, the common data missingness issue challenges prevailing discrete-time approaches, which may yield misleading estimates of lagged effects by assuming fixed measurement intervals. To address these gaps, we conducted an exploratory continuous-time structural equation modeling (CT-SEM) analysis. We used data from the Digital Clinic, an 8-week hybrid care model integrating CBT-based telehealth with smartphone-based digital phenotyping. We fit a continuous-time vector autoregressive model to map daily dynamic relationships among passive sensing (hometime, step count, screen time) and active EMA measures (mood symptom severity, functional impairment, and sleep duration). Responder status (end-of-treatment PHQ-9 < 5 or ≥50% symptom reduction) was included as a time-invariant moderator of the drift matrix to examine differential network dynamics by treatment outcome. Across 218 participants with mild depression or anxiety symptoms (5,844 observations, responder status moderated 13/36 drift parameters. Most notably, responders displayed markedly faster autoregressive self-regulation of sleep duration (ΔAR = −0.85; responder AR = −0.79 vs. non-responder AR = +0.06) and functional impairment (ΔAR = −0.40), with near-zero sleep AR in non-responders suggesting sleep duration instability as a potential marker of poor treatment response. Responders additionally showed stronger screen-tostep (Δβ = +0.82) and step-to-sleep duration (Δβ = +0.60) coupling, a tighter sleep duration– mood symptom severity link (Δβ = +0.16), and stronger negative coupling between functional impairment and hometime (Δβ = −0.40), suggesting that when responders experience functional difficulty they are less likely to withdraw to the home environment. By contrast, non-responders showed a positive effect of hometime on mood symptom severity (β = +0.30 vs. β ≈ 0.00 in Responders), consistent with a social isolation–mood deterioration dynamic absent in treatment responders. Responders further showed lower long-run equilibrium levels across five behavioral and symptom dimensions (hometime, screen time, mood symptom severity, functional impairment, and sleep duration), reflecting systematically altered behavioral baselines associated with treatment response. Poster Session II with Lunch