Abstract

Early warning systems grounded in learning analytics offer a practical route to identifying students at risk of failing, yet their adoption in Chilean undergraduate STEM courses is constrained by limited infrastructure and data availability. This study proposes a parsimonious methodology based on univariate logistic regression with the first partial grade as the sole predictor-a deliberate choice prioritising interpretability, actionable timing, and replicability without specialised analytical infrastructure. The approach was tested on two introductory mathematics courses (Linear Algebra and Mathematics I) at two Chilean universities across 2022 and 2023 cohorts. Models trained on 2022 data were externally validated on 2023 data without retraining. Discrimination was high at both institutions (AUC 0.85-0.93 in external validation). At the institution where approval prevalence remained stable across cohorts, operational metrics (PPV, NPV, balanced accuracy) were consistent between years. At the institution where prevalence shifted markedly-because the second semester is composed mainly of students who previously failed-the model preserved its discriminatory capacity but PPV and NPV inverted substantially, illustrating that operational metrics are functions of prevalence rather than indicators of model failure. A gender equity audit revealed systematically lower sensitivity for female students, indicating under-identification of at-risk women under a uniform threshold and underscoring the need for periodic equity monitoring in any deployment. The methodology offers a replicable, low-overhead floor for early warning practice in comparable courses.