This page was translated automatically from Turkish. Read the original

This page is an AI-assisted Turkish summary of the study presented at the 26th World Congress of Psychiatry. The study belongs to the presenting authors; please refer to the original source when citing it.

EV0446

Same diagnosis, different care: personalizing depression care pathways to reflect individual differences

SAME DIAGNOSIS, DIFFERENT CARE: PERSONALIZING DEPRESSION CARE PATHWAYS TO REFLECT INDIVIDUAL DIFFERENCES

A model combining symptom-level machine learning with treatment dropout risk predicts greater symptom improvement than the assigned treatment.

The aim is to develop a personalized treatment recommendation framework that accounts for symptom–treatment interactions overlooked by models based on total depression severity scores. As its rationale, the poster notes that treatment disengagement affects approximately 20% to 30% of patients and is associated with a two- to fourfold increase in the risk of suicide attempts, yet is often excluded from analyses as missing data. The methods compared population-average modeling with symptom-level LightGBM modeling (gradient-boosted decision trees) using an input matrix of baseline symptom features; a response and remission classifier was developed for SSRIs, a symptom improvement regression model for other antidepressants, and a disengagement classifier for the combination arm. Model performance was evaluated using AUC, RMSE, a paired t-test, and R-squared. In the symptom-level LightGBM model, agreement between predicted and observed PHQ-9 symptom improvement was reported as r = 0.92 and RMSE = 3.86. A paired comparison of the treatment assigned within STAR*D and the model-recommended treatment predicted greater symptom improvement under the model recommendation (t = 4.27, p < 0.001). Implementing the framework is projected to save $8 million per 1,000 patients and produce an improvement of more than 1 point on the PHQ-9 in 570 of 1,000 patients. The authors state that this is the first machine learning framework to integrate symptom-level heterogeneity with treatment dropout risk, and propose validation in independent clinical trial and real-world electronic health record datasets, along with more robust causal inference methods such as doubly robust estimation, as next steps.

All WCP 2026 summaries