IDENTIFYING AND PREDICTING SYMPTOM TRAJECTORIES IN A BLENDED HYBRID DIGITAL MENTAL HEALTH CLINIC

Objective

Blended hybrid digital mental health care is a rapidly expanding treatment model, yet heterogeneity in patient outcomes within these settings remains poorly characterized. We applied latent growth mixture modeling to identify distinct trajectories of depression and anxiety symptoms in a digital psychiatry clinic and examined which digital phenotyping features best discriminate trajectory class membership.

Methods

We analyzed weekly PHQ-9 and GAD-7 scores from 249 patients receiving care in a blended hybrid clinic over an 8-week treatment period. Latent growth mixture models with one to five classes were compared using AIC, BIC, and CAIC to determine the optimal trajectory structure. Following trajectory identification, multinomial logistic regression with 5-fold crossvalidation was used to evaluate the ability of changes in digital phenotyping measures to discriminate between trajectory classes. Model performance was assessed using one-vs-all receiver operating characteristic (ROC) analysis with macro-averaged area under the curve (AUC).

Results

For anxiety (GAD-7), three trajectories were identified: Fast Responders (16.1%, n=40), Responders (58.2%, n=145), and Non-Responders (25.7%, n=64). Similarly, three depression (PHQ-9) trajectories emerged: Fast Responders (5.6%, n=14), Responders (75.5%, n=188), and Non-Responders (18.9%, n=47). Multinomial classification models using digital phenotyping features showed the strongest discrimination for Fast Responders (AUC=0.77 for PHQ-9; AUC=0.80 for GAD-7). Performance was moderate for Responders (AUC=0.60 and 0.62, respectively) and poor for Non-Responders (AUC=0.58 and 0.37). Macro-averaged AUCs were 0.65 for depression and 0.60 for anxiety trajectory classification.

Conclusions

Treatment response in blended hybrid care is highly heterogeneous, with approximately one in four patients with anxiety and one in five patients with depression showing minimal improvement despite completing treatment. Digital phenotyping features reliably identified patients experiencing rapid improvement but struggled to distinguish non-responders from the broader treatment population. These findings motivate further exploration of additional digital biomarkers such as app engagement patterns to improve prospective identification of patients at risk for non-response in digital mental health care.