Cost-Sensitive DASS-42 Risk Classification in Young Adults Using Asymmetric Misclassification Costs

Young adults experience elevated mental health risk, yet most DASS-42 machine learning studies use heterogeneous populations, single-split validation, and accuracy-driven evaluation that treats clinically harmful errors as equal. This study developed a cost-sensitive machine learning framework for three-class mental health risk screening among young adults aged 18–25. A total of 23,039 DASS-42 records were filtered from an international survey and labeled as Low, Moderate, or High risk using the Maximum Severity Rule across depression, anxiety, and stress scores. Six models Naïve Bayes, Decision Tree, Random Forest, XGBoost, LightGBM, and Artificial Neural Network were compared. Class imbalance was handled using class weighting, while Cost-Proportionate Weighting imposed larger penalties on high-risk underclassification through an asymmetric cost matrix. Six models were evaluated using 10-fold stratified cross-validation and held-out testing, with nested hyperparameter tuning for the four tree-based models and fold-wise early stopping for the neural network. Performance was measured using Macro F1-score, ROC-AUC, confusion matrices, and total misclassification cost. The neural network achieved the strongest baseline performance (Macro F1 = 0.9686). Under CPW, its total CSL cost fell to 131, the lowest among all six models. CPW cut cost across all models, from 5.4% to 31.05%, with the neural network showing the largest reduction. Five of six models produced zero Low-High errors; only Decision Tree missed that mark. A rule-based scoring classifier reached 100% accuracy by construction, indicating near-perfect model scores reflect reconstruction of the scoring rule rather than genuine clinical prediction. These findings support cost-aware DASS-42-based screening for early risk identification.