Early warning of academic risk from the first four weeks of online learning activity and assessment performance: a leakage-controlled, demographic-data-minimizing study on the OULAD benchmark

Early warning systems may identify academic risk while support remains possible, but published evaluations often use late-course data, sensitive background information, or leakage-prone splits. We tested prediction from the first four weeks of virtual learning environment activity and assessment performance, quantified the contribution of background/administrative data, and examined transfer across cohorts and modules. Using the Open University Learning Analytics Dataset (32,593 enrollments; analytic n = 27,553 still registered at day 28; 10.7 million activity records), we predicted failure or subsequent withdrawal from 13 behavioral and 9 background/administrative features. Splits and bootstrap resampling were grouped by student, with model selection by grouped cross-validation. The selected in-pool model reached ROC-AUC 0.7747 (student-cluster 95% CI 0.7624–0.7870); behavior alone reached 0.7504, a difference of 0.0243 (95% CI 0.0165–0.0321). A strict availability analysis excluding early grades, whose release date OULAD does not record, retained all-feature AUC 0.7381. At the operational top 5%, precision fell from 0.938 to 0.866 without the background set in-pool and from 0.746 to 0.569 across years. Subgroup disparities persisted after removal; the only gap surviving multiplicity correction was gender for the behavior-only model (0.054 AUC), and removal widened this gap in the main paired analysis. Thus, data minimization is not by itself a fairness remedy. Selection confined to 2013 chose an all-feature random forest that reached AUC 0.7470 (95% CI 0.7393–0.7549) in 2014. Transfer to unseen modules was uneven (0.64–0.79); an exploratory seven-module comparison found the weakest transfer where no assessment fell inside the prediction window, an association that is explicitly non-causal. The findings support a prospective, locally validated pilot of a week-four risk-ranking component without background profiling, provided subgroup performance is audited and calibration is assessed using completed cohorts before each new presentation. Code and outputs make every reported number traceable. This study establishes predictive feasibility, not that alerts or support interventions improve student outcomes.