<![CDATA[The high dropout rate in health science programs not only affects the efficiency of educational management but also threatens the stability of the professional medical workforce. This research proposes a Hierarchical Early Warning System (HEWS) framework using the Light Gradient Boosting Machine (LightGBM) algorithm to predict student dropout risks at an early stage, utilizing a primary dataset of academic and demographic records from an Indonesian health sciences faculty. While previous studies predominantly utilized full-trajectory academic records, this study compares the model's performance across two temporal scenarios: early detection using historical data up to the second semester and full-trajectory evaluation. The Synthetic Minority Over-sampling Technique (SMOTE) is applied to address the extreme class imbalance inherent in educational datasets. Experimental results demonstrate that the proposed model achieves an F1-Score of 0.95 for the minority class in both internal scenarios and maintains a robust F1-Score of 0.91 when validated against an international dataset. Furthermore, SHapley Additive exPlanations (SHAP) is integrated to dissect the algorithm's black-box nature. The SHAP analysis reveals a critical shift in risk drivers: high school education quality and first-semester grades dominate the early detection phase, whereas fourth-semester grades and sick leave frequencies become primary predictors in the full-trajectory phase. These findings provide actionable insights for university management to design targeted academic and psychological interventions, thereby supporting the achievement of Sustainable Development Goals (SDG) 3 and 4.]]>
Copyrights © 2026
