Artificial Intelligence for Predicting Student Dropout from Digital Traces in Higher Education: A Systematic Review of Methods, Explainability, and Fairness Gaps for Equitable and Sustainable Retention

Student dropout limits the equal access to tertiary education sought by Sustainable Development Goal 4, and universities increasingly use machine learning models trained on student digital traces to identify students at risk. Following PRISMA 2020, this systematic review examined how such models are built, explained, and reported. Peer-reviewed empirical studies published in 2015–2025 were retrieved from Scopus and Web of Science, excluding online-only education. After re-assessment with written rules and independent double coding, 19 reports were classified by prediction target, horizon, data source, model family, explainability, and validation, and audited for reporting completeness with 16 items adapted from TRIPOD+AI and for signals of analytical quality. The main finding concerns the target: although framed as dropout prediction, 15 reports predicted course failure, so the evidence concerns academic risk more than departure. Learning management system records appeared in 16, ten compared machine and deep learning in parallel, and 11 used explainability, mainly to describe global risk factors. Ethical approval was documented in 8 reports, none reported performance across protected subgroups, and 16 lacked temporal or external validation. These gaps concern what is reported rather than proving biased models, but they prevent institutions from verifying whether early-warning systems serve all students equally.