Early Warning, Grade Prediction, and Teacher-Facing LLM-Ready Explanations Toward an Open Volleyball Course: Reproducible Evidence from Four Public Education Datasets

Open online courses and public skill-development programs often experience learner dropout not due to content limitations, but because instructors receive delayed and non-actionable feedback. This study proposes and empirically evaluates an integrated framework for an open volleyball course that combines early warning prediction, grade estimation, and teacher-oriented LLM-generated academic status explanations. The predictive models were tested on four lightweight educational datasets: xAPI-Edu-Data, Predict Students’ Dropout and Academic Success, Student Performance, and Higher Education Students Performance Evaluation. A unified preprocessing pipeline was applied using one-hot encoding, an 80/20 train-test split, and 5-fold cross-validation. Decision Tree, Random Forest, and XGBoost models were evaluated for classification, alongside their regression variants for grade prediction. Results show consistent performance across datasets. Random Forest achieved the best macro-F1 on xAPI-Edu-Data (0.799) with a macro-AUC of 0.914, while XGBoost performed best on the dropout dataset (macro-F1 = 0.689, macro-AUC = 0.892). For Student Performance, early-warning models without prior grades reached an RMSE of 3.086, improving to 1.398 when full information was available. On the higher education dataset, performance remained limited due to small sample size and multi-grade targets, with Random Forest achieving a macro-F1 of 0.248. Ablation results confirmed that behavioral and progression features significantly improve predictive accuracy. An explanation layer translated model outputs into structured, teacher-facing natural language linked to key risk indicators and intervention cues. Overall, the framework demonstrates analytic feasibility for structured volleyball course monitoring, though results should be interpreted as pre-deployment evidence rather than validation in real instructional settings. Explanation quality improves when grounded in observed behavioral signals rather than generic generation.