A Cognitive Behavioral Therapy Informed Explainable AI Framework for Academic Performance Analysis

Academic performance prediction models often prioritize accuracy over interpretability, limiting their practical utility in educational interventions. This study integrates Cognitive Behavioral Therapy (CBT) principles with explainable artificial intelligence (XAI) to analyze academic performance through behavioral, emotional, and cognitive dimensions. A dataset of 1,000 student records was analyzed using exploratory data analysis, multi-class classification (Logistic Regression, SVM, Random Forest), regression modeling (Linear Regression, Random Forest Regressor, SVR), and SHAP-based explainability analysis. Models were evaluated with and without cognitive proxy variables to assess the independent contribution of behavioral and emotional factors. Classification models achieved near-perfect accuracy (Random Forest: 1.00, SVM: 0.945, Logistic Regression: 0.98) when prior academic score was included, indicating proxy leakage. After removing this cognitive proxy, accuracy dropped dramatically to 0.315 (0.325 with SMOTE oversampling). Regression models predicting academic score from behavioral features yielded negative R² values (Linear: −0.01, Random Forest: −0.05, SVR: −0.17), with RMSE exceeding 18 points. SHAP analysis revealed weak and unstable feature contributions among behavioral variables, with interaction values centered near zero. The findings demonstrate that behavioral and emotional variables alone lack sufficient predictive power for academic performance modeling, supporting CBT theory that cognition mediates the relationship between behavior and outcomes. This study emphasizes the critical role of explainable AI in identifying model limitations and theoretical inconsistencies, advocating for cognitive construct inclusion in educational prediction models.