Actionable and Trustworthy AI for Student Depression Prediction Using Multi-Level Explainability
Abstract
It is critical to be able to identify and intervene with high-risk students as early as possible - this is most easily accomplished by developing scalable systems for providing such early warning signals. Although many researchers have developed highly accurate machine learning (ML) models to detect depression, poor clinical transparency makes it challenging to get adoption from both educators and physicians. This study presents a socio-technical framework addressing the ethical and clinical issues arising from using 'black-box' predictive models for high-stakes student mental health screening within computational psychiatry. By applying a two-level stacking ensemble that combines Random Forest (RF), Support Vector Machine (SVM) and Logistic Regression (LR) as base learners with a LR meta-learner for optimizing predictive performance. Class imbalance typical of mental health data is mitigated using Synthetic Minority Over-sampling Technique (SMOTE). Embedded within this socio-technical system, a dual-layered Explainable Artificial Intelligence (XAI) subsystem of SHAP and DiCE explains the model's inductive biases and proposes individually actionable recourse to high-risk students, ensuring ethical clinical translation. Empirically, the model exhibits 0.8421 accuracy and 0.9184 ROC-AUC. This study contributes to the deployment of interpretable ML within computational psychiatry and presents an extensible ethical framework that offers both accurate mental health risk identification and individualizable mental health intervention for university students.