AN INTERPRETABLE MACHINE LEARNING FRAMEWORK FOR PREDICTING STUDENT BURNOUT RISK LEVEL IN AI-ASSISTED LEARNING ENVIRONMENTS

Abstract

The rapid adoption of generative AI (Gen AI) tools in higher education has raised concerns about their role in student burnout. However, most predictive models linking AI usage to wellbeing outcomes operate as black boxes, limiting their practical application by educators and institutions. This study presents an interpretable machine learning (ML) framework for predicting student burnout risk in AI-assisted learning environment. It combines supervised classification with SHAP-based explainability to determine which AI-usage and traditional academic behaviors drive risk. We trained and compared RF, DT, LR, XGBoost, and LightGBM classifiers on a dataset of 50,000 students, which included Gen AI usage patterns, prompt engineering skill and traditional study behaviors, as well as a burnout label. The top-performing model, LightGBM, had an F1-score of 0.639 and a ROC-AUC score of 0.717. SHAP-based interpretability revealed that Gen AI dependency per amount of overall study time was one of the top predictors of burnout risk, implying that subjective reliance on AI tools carries a stronger predictive signal than overall time spent using only these tools. These findings suggest that interpretable ML frameworks can help with the early identification of at-risk students in AI-integrated learning environments, as well as provide institutions with actionable feature insights to guide interventions and policies around Gen AI use.