Multi-Label Prediction of Adolescent Mental Health Comorbidity based on Psychosocial Factors using XGBoost and Explainable AI

Abstract

Adolescent mental health is an important issue that can affect quality of life and academic performance. Screening approaches based on the Depression, Anxiety, and Stress Scales (DASS-42) typically predict each condition separately, making them less capable of representing the comorbidities that frequently occur among depression, anxiety, and stress. This study aims to develop a multi-label classification model for predicting adolescent mental health risks that can capture inter-label dependencies while addressing class imbalance. The proposed model integrates eXtreme Gradient Boosting (XGBoost) with a Classifier Chains architecture, while data imbalance is addressed using the Multi-Label Synthetic Minority Over-sampling Technique (MLSMOTE). The dataset was obtained from the DASS-19 repository on the Kaggle platform, based on the DASS-42 questionnaire instrument, and underwent preprocessing, resulting in 10,315 respondents with 15 psychosocial features as predictors. Experimental results show that the XGBoost Classifier Chains model achieved the best performance, with a Macro F1-Score of 0.8774, Hamming Loss of 0.1894, and Exact Match of 0.6413 on the test data. Analysis using SHapley Additive exPlanations (SHAP) revealed that several personality-related features made dominant contributions to the predictions and confirmed the presence of inter-label dependencies effectively exploited by the Classifier Chains architecture. These findings indicate that the proposed approach has the potential to support transparent machine learning-based early screening and assist in identifying mental health risks among adolescents.