Early Detection of Comorbid Anxiety and Depression Using Explainable Machine Learning on DASS-21 Psychometric Data

Abstract: Depression and anxiety frequently co-occur, yet early detection of their comorbidity remains challenging due to reliance on subjective clinical assessments. This study presents an explainable machine learning framework for binary classification of joint anxiety-depression at-risk status using the DASS-21 psychometric questionnaire. To prevent data leakage, we employ a stress-proxy feature strategy that excludes depression and anxiety subscale items from the input features, retaining only stress-related questionnaire items and demographic variables. Six classifiers—Logistic Regression, SVM, Random Forest, XGBoost, Gradient Boosting, and an MLP neural network—are evaluated using 5- fold stratified cross-validation with SMOTE-based class balancing. The best-performing model, Random Forest (tuned), achieves a test accuracy of 60.10% and ROC-AUC of 0.4917 under the leakage-free setting, highlighting the inherent difficulty of predicting comorbid risk from indirect indicators alone. SHAP (SHapley Additive exPlanations) analysis identifies education level and DASS-21 item Q1A (difficulty winding down) as the most influential predictors. Demographic fairness analysis reveals comparable performance across gender and age subgroups. These findings establish a transparent, reproducible baseline for comorbid mental health screening and underscore the need for richer multi-modal feature sets to improve predictive accuracy.

Keywords: machine learning; explainable AI; depression; anxiety; comorbidity; DASS-21; SHAP; mental health screening