Advancing patient stratification in major depressive disorder: Evaluation of clinical staging and machine learning prediction models based on real-world data

OBJECTIVE: To evaluate the utility of a clinical staging model and compared its prognostic performance with an unsupervised machine learning-based stratification method in a real-world cohort of patients with Major Depressive Disorder (MDD). METHODS: A longitudinal, observational study was conducted at a public mental health institution. 262 adult outpatients with MDD were classified at baseline using the Hetrick and McGorry staging model. Multivariate multinomial regression identified baseline correlates of stage, and generalized linear models (GLMs) assessed the model's predictive validity on symptom severity (CGI, PHQ-9) at 9-month follow-up (n = 198). Additionally, k-means clustering was performed to identify clinically meaningful subgroups based on key baseline clinical variables (illness duration, diagnosis, comorbidity, CGI, PHQ-9 and GAF), and its predictive performance was compared to the clinical staging framework. RESULTS: Multivariate analysis showed longer illness duration and psychiatric comorbidities were robust predictors of advanced stages. However, clinical staging explained a small proportion of variance in follow-up outcomes (2 % for CGI and less than 1 % for PHQ-9). Patients in advanced clinical stages showed significantly worse outcomes, although effect sizes were small. Clustering identified three subgroups, with partial overlap with clinical stages. Cluster membership showed a comparatively stronger, though still modest, association with 9-month outcomes, with clusters 2 and 3 showing significant associations with increased CGI (β = 0.9-1.3) and PHQ-9 scores (β = 3.9-6.9). CONCLUSION: The explanatory power of clinical staging was limited. Data-driven clustering offered modestly superior prognostic value and may complement staging approaches to enhance patient stratification in MDD.