Artificial intelligence-driven early warning for online learner dropout: temporal explainability and theory-grounded intervention

Online learning faces persistently high dropout rates, yet most prediction models are static, single-time-point classifiers with little interpretability and untested cross-course transfer. Drawing on 32,593 registrations from the Open University Learning Analytics Dataset (OULAD), we build a temporally dynamic, interpretable machine learning framework. Its three-layer architecture spans static entry attributes, dynamic behavioral indicators and trend measures, each mapped to a construct from self-regulated learning, Community of Inquiry or integration theory. Six algorithms are benchmarked across five cumulative observation windows. LightGBM is best or joint-best at every window, reaching AUC 0.727 before the course opens and 0.911 by day 120. Temporal SHAP analysis identifies the engagement persistence ratio as the leading predictor from mid-course onward. A collapse of activity in the second half of the window raises risk, whereas sustained activity lowers it. The six trend features alone carry 25.4% of the total attribution. Leave-One-Course-Out cross-validation returns 0.857, and a prospective analysis restricted to learners still enrolled returns 0.72 to 0.75, the figure an institution should plan against. In conclusion, the attribution results can be translated into a four-tier protocol that matches an intervention to each learner’s risk factors, so a fixed outreach budget reaches failing engagement months before withdrawal.