Enhancing students’ academic outcomes in blended learning (BL) environments requires timely intervention for at-risk students, making early detection a critical task for educators and researchers. However, this task remains challenging due to imbalanced educational data and limited model sensitivity. This study investigated the application of machine learning (ML) for the early identification of at-risk students in BL environments, aiming to improve educational quality in alignment with Sustainable Development Goal 4 (SDG 4). A dataset of 342 students, collected from a BL environment at the Institute of Technology of Cambodia (ITC), was used to model academic performance as a three‑class classification problem: At‑risk, Good, and Excellent. To address class imbalance, the Synthetic Minority Over‑sampling Technique (SMOTE) was employed, and multiple models including Random Forest, Gradient Boosting, and ensemble methods were evaluated using stratified cross‑validation. The best multi‑class performance was achieved by Gradient Boosting with SMOTE, yielding an accuracy of 75.43% and a macro F1‑score of 68.74%. Nevertheless, this model detected at‑risk students only 41.67% of the time. To resolve this limitation, a binary classification approach with a focus on recall was proposed. By reframing the problem as an early‑warning task and lowering the classification threshold from 0.50 to 0.10, recall improve dramatically from 52.78% to 91.67%, allowing for highly sensitive detection of at‑risk students. However, this improvement was at the cost of lower precision with more false positives. To overcome this trade-off, we further evaluated the F2-score and threshold analysis and identified a threshold of around 0.40 as a better trade-off between recall and precision in real educational intervention scenarios.  These results indicate that threshold optimization is an essential technique for educational early‑warning systems, where recall is more important than overall accuracy. The study offers a practical and scalable framework for improving student monitoring, enabling timely intervention, and enhancing learning outcomes in blended learning environments.