Answers are grounded in this paper and cite their sources.
University preparatory and foundation programmes often experience higher attrition than standard degree pathways, yet instructors often lack reliable early indicators of students beginning to disengage. This paper presents EduRiskNet, a calibrated, explainable early-warning pipeline designed to identify at-risk students in preparatory information-technology (IT) courses. In this study, we validate the pipeline on a public benchmark; we do not analyse institutional IT-course data. The pipeline uses learning-management-system/virtual-learning-environment activity, assessment behaviour, and static student-profile information to construct as-of-checkpoint features (using only information available at the prediction point) at successive weekly checkpoints. Using the public Open University Learning Analytics Dataset (OULAD), we evaluate logistic regression, random forest, histogram gradient boosting, and a feed-forward multilayer perceptron (MLP) using a time-respecting split in which the 2014J presentation is held out as an unseen test cohort. EduRiskNet denotes the complete calibrated pipeline rather than a particular classifier; the random forest is its operative model. The OULAD benchmark contains 32,593 student registrations, 22 module presentations, and over 10.6 million VLE interaction records. The results show that early prediction of at-risk students is feasible. At week 4, the random forest performed best overall, with an ROC-AUC of 82.6% (95% CI 81.8–83.4%), an AUPRC of 86.0%, an at-risk recall of 72.0%, and an expected calibration error of 0.031. At operational dashboard checkpoints, performance improved from ROC-AUC 81.4% and F1 72.4% at week 3 to ROC-AUC 89.1% and F1 80.6% at week 10. At the week-10 checkpoint, the tiering layer of a conceptual instructor dashboard assigned the 11,260 students of the 2014J test cohort to high (4188; 37.2%), medium (2247; 20.0%), and low (4825; 42.9%) risk tiers. Compared with recent OULAD studies, the week-4 result compared favourably with a leakage-controlled day-28 gradient-boosting benchmark (ROC-AUC 79.0%), although differences in cohort definition and data splitting limit direct comparison. The strongest predictors were early assessment performance, LMS activity consistency, submission behaviour, prior education, and missed early assignments. A gender subgroup audit of the MLP baseline revealed statistically clear differences in recall, false-positive rate, and ROC-AUC between female and male students, highlighting the need for human-in-the-loop review and responsible deployment. The study contributes a reproducible, calibrated, and explainable benchmark pipeline for early-warning learning analytics, while noting that validation on preparatory IT-course data, instructor usability, and intervention effects remain to be established empirically.
No takes yet. Share an insight, caveat, or question.
Venugopal et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: