Pre-usage Perceptual Signals as Early Warning Indicators: A Machine Learning Framework for At-risk Student Detection in the Educational Metaverse

Immersive learning environments, particularly the educational metaverse, offer substantial opportunities for interactive, collaborative, and personalised learning. However, many students struggle during initial adoption, often leading to disengagement or dropout. This paper introduces and evaluates, at a proof-of-concept level, a perceptual prediction framework for the early identification of students showing low adoption propensity (a pre-usage risk marker for subsequent disengagement) in educational metaverse environments, relying exclusively on perceptual data. Distinct from traditional models, this framework purposefully excludes Behavioural Intention (BI) from the predictor set, using it solely to operationalise the risk target, and focuses instead on the instrumental and affective appraisals that precede explicit intent. The framework integrates seven perceptual dimensions drawn from the Technology Acceptance Model and its affective and trust-related extensions, including Hedonic Motivation, Perceived Enjoyment, Perceived Usefulness and Attitude. Analyses draw on a publicly available secondary dataset of 215 Malaysian university students, collected through a single cross-sectional questionnaire administration prior to any metaverse exposure. K-means clustering yields four engagement readiness strata (retained on pedagogical rather than statistical grounds; internal validity indices favour a two-cluster solution), with a low-readiness stratum ( $$n = 31$$ ; 14.4%) constituting the primary intervention target. Random Forest and Logistic Regression perform equivalently on the held-out test set (81.40% and 79.07%; exact McNemar test, $$p = 1.00$$ ) and show strong discrimination (macro ROC-AUC =.922 and.939). Because the risk labels are derived from Behavioural Intention, which correlates strongly with all predictors, these results should be read as evidence of perceptual discriminability rather than as an independently validated predictive system. They nonetheless indicate that pre-usage perceptual signals carry sufficient discriminative information to support proactive pedagogical intervention before platform entry. Two constraints qualify these findings: the risk labels derive from a construct that shares substantial variance with the predictors, and heterotrait–monotrait ratios indicate limited discriminant validity among the affective constructs. Future work should establish convergent validity against longitudinal behavioural traces and replicate the framework across culturally and institutionally diverse settings.