From Crisis to Prevention: Machine Learning Reveals Behavioral Predictors Outperform GPA in Early Student Risk Detection

Globally student attrition continues to challenge higher education institutes, with dropout cost estimated at three to five times the investment needed to retain students through targeted support.This study developed and validated an early warning system to identify at-risk undergraduate students within the first six weeks of semester, before traditional grade-based metrics signal problems.Based on institutional data for 154 students collected between 2022 and 2024, a combination of logistic regression and machine learning methods used (specifically, Random Forest and XGBoost) to evaluate which early academic warning signs most accurately predict whether a student will be classified as failing by the end of the semester.Our analysis reveals that behavioral patterns, particularly attendance drops and repeated course withdrawals, substantially outperform GPA decline in early prediction.The logistic regression model achieved outstanding discrimination (AUC = 0.969, 95% CI: 0.946-0.968),validated through cross-validation (mean AUC = 0.935).Feature importance analysis confirmed attendance decline as the dominant predictor, which demonstrates an important score 5.4 times higher than GPA drop.This approach transforms institutional support by introducing a preventive, data-driven model.By acting on early indicators and framework, institutions can implement interventions 4-6 weeks in advance, moving their support system from a reactive stance to one of proactive care.academic crisis, supporting the shift from reactive response to preventative care.