Early Student Risk Detection Using CR-NODE: A Completion-Focused Temporal Approach with Explainable AI
Abdelkarim Bettahi, Hamid Harroud, Fatima-Zahra Belouadha · Algorithms · 2025
Student dropout prediction remains critical in higher education, where timely identification enables effective interventions. Learning Management Systems (LMSs) capture rich temporal data reflecting student behavioral evolution, yet existing approaches underutilize this sequential information. Traditional machine learning methods aggregate behavioral data into static features, discarding dynamic patterns that distinguish successful from at-risk students. While Long Short-Term Memory (LSTM) networks model sequences, they assume discrete time steps and struggle with irregular LMS observation intervals. To address these limitations, we introduce Completion-aware Risk Neural Ordinary Differential Equations (CR-NODE), integrating continuous-time dynamics with completion-focused features for early dropout prediction. CR-NODE employs Neural ODEs to model student behavioral evolution through continuous differential equations, naturally accommodating irregular observation patterns. Additionally, we engineer three completion-focused features: completion rate, early warning score, and engagement variability, derived from root cause analysis. Evaluated on Canvas LMS data from 100,878 enrollments across 89,734 temporal sequences, CR-NODE achieves Macro F1 of 0.8747, significantly outperforming LSTM (0.8123), Extreme Gradient Boosting (XGBoost) (0.8300), and basic Neural ODE (0.8682). McNemar’s test confirms statistical significance (p<0.0001). Cross-dataset validation on the Open University Learning Analytics Dataset (OULAD) demonstrates generalizability, achieving 84.44% accuracy versus state-of-the-art LSTM (83.41%). To support transparent decision-making, SHapley Additive exPlanations (SHAP) analysis reveals completion patterns as the primary prediction drivers.