Can a mobile app improve prognostic accuracy in first episode psychosis? Evidence from ThinkApp using a machine...

Mobile app-based interventions for first-episode psychosis (FEP) can extend care, yet heterogenous engagement obscures which components drive benefit. Machine learning models integrating clinical variables with app-usage data may clarify trajectories and guide design. The aims of the study were to identify user profiles and usage patterns in a mobile intervention for youth with FEP (ThinkApp) and test whether adding usage features improves 6-month outcome prediction.

Baseline clinical and sociodemographic variables were merged with engagement metrics. Clustering methods were applied to usage data to explore participation profiles. We trained regression models on baseline-only and enriched (baseline + app) feature sets using Huber, Random Forest, and RANSAC with standardized preprocessing, feature selection and cross-validated evaluation.

We analyzed 26 participants (mean age = 22.60). Three clusters emerged; overall usage time most strongly mapped to cluster membership, and psychoeducation activity was associated with engagement (r = -0.82). Baseline age correlated inversely with login frequency (r = -0.53) but not with total usage time. Enriching models with ThinkApp features improved 6-month predictions, with the largest gain observed for PANSS Positive Scale (ΔR² = 0.085). The number of psychoeducation videos viewed was the most influential predictor.

Exploratory analyses suggested that app-engagement metrics, particularly psychoeducation, may add information to the prediction of 6-month outcomes beyond baseline clinical and sociodemographic variables. Age was linked to access patterns, highlighting the need to tailor digital implementation in early psychosis. These findings require replication in larger samples before their clinical utility can be established.