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.

SO384

Combining age-stratified modeling with explainable artificial intelligence to predict treatment resistance and suicide risk in depression

INTEGRATING AGE-STRATIFIED MODELING AND EXPLAINABLE AI FOR PREDICTING TREATMENT REFRACTORINESS AND SUICIDE RISK IN DEPRESSION

Age-stratified TabNet models using FDG-PET data predicted treatment-resistant depression and suicide risk more accurately than conventional machine learning methods and identified distinct metabolic patterns in each age group.

Based on the premise that the brain mechanisms underlying treatment-resistant depression (TRD) vary across age groups, the study aimed to evaluate whether age-stratified modeling improves prediction by capturing age-specific metabolic patterns. A total of 245 patients with major depressive disorder (aged 21–82 years) from Taipei Veterans General Hospital were divided into three groups: Young (under 25 years; 7 non-treatment-resistant, 17 TRD), Adult (25–49 years; 26 non-treatment-resistant, 58 TRD), and Older (50 years and above; 56 non-treatment-resistant, 81 TRD). Resting-state FDG-PET data were obtained from 116 AAL brain regions, and SUV correction and min–max normalization were applied. An age-stratified framework built using TabNet, a deep learning architecture for tabular data, was compared with conventional machine learning models (SVM, RF, XGBoost). SMOTE, Mixup, and a class-balanced loss function were used to address class imbalance, while model interpretability was provided by SHAP (SHapley Additive exPlanations). TabNet outperformed conventional models across all metrics; accuracy in predicting treatment resistance was 75.33% in the Young group, 71.62% in the Adult group, and 77.80% in the Older group. Accuracy in predicting suicide risk was reported as 86.00% in the Young group, 80.19% in the Adult group, and 70.36% in the Older group. SHAP analysis revealed age-specific neurobiological profiles: differences in sensory integration and social networks (parietal lobe, temporal pole) in young patients; dysregulation among the prefrontal cortex, anterior cingulate cortex, and limbic system in adults; and subcortical and cerebellar involvement (putamen, cerebellum) in the older group. The authors state that age is a key moderator of TRD presentation and model performance, and that stratifying patients by age and using interpretable artificial intelligence such as TabNet both improves predictive accuracy and enables the identification of age-specific pathological mechanisms.

All WCP 2026 summaries