A Privacy-Conscious AI Framework for Early Identification of At-Risk Students Across Disciplines Using LMS Engagement Data

This study explores the application of artificial intelligence (AI) in higher education to enable the early identification of at-risk students using only engagement data from learning management systems (LMS). Unlike many existing early-warning models that are limited to single disciplines or rely on sensitive demographic and prior academic records, the proposed approach offers a privacy-conscious and highly generalizable predictive framework suitable for diverse higher education contexts. The dataset includes over 1.7 million LMS interaction records from 236 undergraduate subjects spanning four academic divisions. These subjects encompass a wide variety of instructional designs and assessment structures. To address cross-subject heterogeneity, this study employs rank-based engagement features that represent students’ relative behavioral patterns within each course, facilitating meaningful comparison across disciplines without reliance on absolute activity levels. Using standard machine learning classifiers, the model achieves over 90% prediction accuracy for final subject performance by Week 3 of the semester, demonstrating that reliable early detection of at-risk students is feasible at an early stage of teaching and learning. Rather than claiming intervention effectiveness, the study positions AI-enabled early prediction as a scalable foundation for proactive student support and enhanced teaching responsiveness, with the potential to inform timely pedagogical actions such as targeted outreach and academic advising. By emphasizing generalizability, ethical data use, and privacy protection in AI-enabled learning analytics, this research contributes practical insights into how predictive AI can responsibly support teaching and learning in higher education.