Machine Learning-Based Early Warning for Student Dropout: Evidence from LMS Behavioral Engagement Patterns in Online Higher Education

Abstract

Student dropout in online higher education remains critically high, far exceeding face-to-face rates, yet declining behavioral activity in Learning Management Systems (LMS) offers key signals for early intervention. To identify robust predictors and model suitability for Early Warning Systems (EWS), this study presents a comparative analysis of machine learning for dropout prediction using clickstream data from the Open University Learning Analytics Dataset (OULAD), covering 32,593 students across seven undergraduate modules. Three supervised algorithms with Logistic Regression, Random Forest, and Support Vector Machine (SVM), were trained on 13 engineered features combining behavioral and demographic attributes from the Virtual Learning Environment (VLE), with Recall prioritized to minimize missed at-risk students. Results demonstrate that all models achieved strong discriminatory performance with AUC-ROC > 0.93; specifically, SVM provided the highest EWS fit with recall of 0.903, missing only 196 of 2,031 withdrawals (9.7%), while Random Forest attained the best overall accuracy (0.866) and AUC-ROC (0.940). Feature importance analysis further revealed that VLE behavior accounted for 85.0% of predictive power, with Activity Span emerging as the dominant predictor at 41.3%. Cross-module validation confirmed temporal engagement consistency as a robust, generalizable dropout signal. Therefore, these findings provide practical guidance for implementing data-driven EWS in online learning by prioritizing behavioral span metrics over static demographics