Explainable AI for Early Detection of Academic Disengagement in Blended Learning Environments

Abstract

Learning management systems generate detailed behavioural traces, and predictive models built on those traces can identify students at risk of failure or withdrawal well before conventional assessment does. As these models have moved from research prototypes toward institutional deployment, opacity has become the binding constraint: an instructor asked to act on an unexplained risk flag has no basis for choosing an intervention, and no basis for deciding whether the flag is wrong. Explainable artificial intelligence has been proposed as the remedy, with feature attribution methods now standard in the learning analytics literature. This review examines that proposal critically. It argues that the field has adopted attribution methods to satisfy a legitimate demand for transparency, but has largely not asked whether the explanations produced are of the kind instructors actually need. Three problems are developed. Disengagement is treated as equivalent to its behavioural proxies, so models explain platform inactivity rather than the underlying construct, and blended delivery makes this substitution especially unsafe because much learning occurs off-platform. Attribution explains the model rather than the student, and the features that carry the most attribution are frequently not the features an instructor can act on. And the evaluation literature measures predictive accuracy while the outcome that matters — whether flagged students are helped — is rarely measured at all. The paper proposes reorientation toward actionable and contestable explanation, evaluation against intervention outcomes, and explicit treatment of the equity and self-fulfilling-prophecy risks that early warning systems carry.