Deep learning-based AI model for predicting academic success and engagement among physical higher education students

Academic achievement and student engagement are essential components of educational success, particularly in physical education (PE) programs, where cognitive and physical competencies intersect. This study introduces HybridStackNet, a stacked ensemble model integrating Random Forest and Support Vector Machine (SVM) as base learners and Logistic Regression as a meta-learner, aimed at jointly predicting academic success and engagement among higher education PE students. The model was trained on a publicly available dataset from Kaggle comprising 500 instances with academic, behavioral, and physical attributes. Preprocessing steps included label encoding, z-score normalization, Random Forest-based feature selection, and SMOTE for class balancing. Using stratified 5-fold cross-validation and GridSearchCV, HybridStackNet demonstrated promising performance (Accuracy = 0.992, Precision = 0.9922, Recall = 0.992, F1-score = 0.9915, AUC = 0.9942, Jaccard = 0.9842, Kappa = 0.9846, and Hamming Loss = 0.008). These results surpassed several baseline models, including Decision Tree, SVM, KNN, Random Forest, and Gradient Boosting. Explainability was explored using Partial Dependence Plots (PDPs) and LIME. PDPs highlighted key feature impacts (e.g., Attendance_Rate, Overall_PE_Performance_Score, Motivation_Level), while LIME provided interpretable thresholds (e.g., Attendance_Rate [Formula: see text], Speed_Agility_Score [Formula: see text]) for local explanations. HybridStackNet and the accompanying explainability framework offer an interpretable machine learning approach to early performance risk detection in PE education settings.