Predicting mental health treatment outcomes using latent growth mixture models and machine learning in a real-world clinical setting.

Abstract

Previous models of depression outcomes have been limited by symptom heterogeneity within populations. This study conducted a retrospective analysis using latent growth mixture models to identify heterogeneous trajectories within a clinical population, subsequently developing machine learning models to predict clinical outcomes based on baseline characteristics and symptom measures. The study analyzed approximately 15,000 clients aged 18-89 in a real-world clinical setting, treated for up to 180-days between 2015 and 2020. Three distinct groups were identified: Low-Low (stable low scores), High-Low (improving scores), and High-High (stable high scores), representing 56%, 18%, and 26% of the cohort, respectively. Significant differences were observed in baseline factors and outcome assessments across these groups. The machine learning model demonstrated balanced accuracies of 0.67-0.93, and ROC-AUC values between 0.87 and 0.97 in predicting group membership from baseline data. Key predictors included baseline PHQ-9 scores, sex, age, and PTSD diagnosis. Prospective application of the model accurately categorized new client data, aligning predictions with observed outcomes. The study highlights the potential of machine learning in clinical settings to predict client outcomes, offering valuable insights based on initial data, potentially enhancing treatment personalization and efficacy. Further research should explore the impact of individual treatment modalities to validate the model's clinical utility.