Predicting at-risk undergraduate students and high-impact modules: evidence from a South African University of Technology

Scientific problem High failure and dropout rates remain a persistent challenge at South African Universities of Technology, where routinely collected institutional data are seldom used systematically to identify at-risk students or structurally high-impact modules at the faculty level. Aim This study identifies the key predictors of academic risk and determines which modules most negatively influence student progression within the Faculty of Management Sciences at Durban University of Technology. Method Quantitative, non-experimental research design grounded in the post-positivist paradigm and informed by Tinto's model of student departure and Bourdieu's theory of cultural capital. Anonymised institutional records spanning 2017–2025 were analysed using binary logistic regression and random forest classification. A retrospective full-programme model was developed for explanatory purposes, while a deployable Year 1 early-warning model was constructed to support prospective student risk identification. Results The full-programme Student Risk Classification Model achieved an ROC-AUC of 0.888, with sensitivity of 59.0% and specificity of 92.7% at the 0.50 classification threshold. The Year 1 early-warning model achieved an ROC-AUC of 0.835 and identified 78.9% of future non-graduates before Year 2, with a specificity of 74.0% at the Youden-optimal threshold. Academic performance variables were substantially stronger predictors than demographic characteristics, while first-time entering status, race, and gender were not statistically significant within the Diploma-restricted sample. Twenty-two high-impact modules were identified, of which eighteen remained statistically significant after false-discovery-rate correction. These modules were concentrated within three departments, with student-level and module-level risk indicators converging on the same departments, providing multi-level evidence that academic risk is structural rather than compositional. Conclusion Learning analytics can be effectively implemented using routinely collected institutional data without requiring specialist infrastructure. The study contributes two practical decision-support instruments—a Student Risk Classification Model and a Module Impact Index—that provide a replicable faculty-level framework for early intervention, targeted academic support, and evidence-based curriculum reform. Future research Future studies should externally validate the Student Risk Classification Model and Module Impact Index across other faculties and higher education institutions. Incorporating psychosocial, socio-economic, behavioural, and learning management system data may further improve predictive accuracy and strengthen institutional early-warning systems.