Academic risk classification can help universities identify students who require closer academic monitoring, but imbalanced classes and limited model transparency can reduce practical usefulness. Objective: This study develops a Random Forest classifier for academic-risk categories and explains model decisions using SHAP. Methods: The study used academic data from 484 Computer Science students at Universitas Islam Negeri Sumatera Utara for the 2025/2026 odd-to-even academic period. The target was formed using rule-based indicators, while the final predictors were Jumlah_MK (number of courses), Praktikum (practicum courses), and MK_Berat (heavy courses). Data were split stratified into 387 training and 97 testing observations. SMOTE was applied only to training folds during stratified five-fold cross-validation and again to the complete training set for the final model. Results: Cross-validation produced mean Accuracy 91.72%, Precision 46.05%, Recall 81.33%, F1-Score 58.21%, and ROC-AUC 90.46%. On the original testing set, the model achieved Accuracy 97.94%, Precision 85.71%, Recall 85.71%, F1-Score 85.71%, and ROC-AUC 98.65%, with TN=89, FP=1, FN=1, and TP=6. Feature Importance and Mean Absolute SHAP both ranked Jumlah_MK as the most influential predictor. Conclusion: Random Forest with training-only SMOTE provided strong testing performance and improved minority-class detection, while SHAP added an interpretable layer. The results should be interpreted cautiously because the high-risk class was small and the study used a single academic period