Stage Aware Student Dropout Prediction in Higher Education

Student dropout remains a persistent challenge for higher education institutions, and predictive analytics promises earlier and better targeted support. Published studies, however, differ widely in what they call dropout, when they predict it, which data they use and how they evaluate their models. This paper presents a comparative synthesis of four recent studies that span the main design choices in the field: a class level platform built on attendance data at the University of Coimbra; a multi stage cascade of models built on academic data only for Spanish engineering students; a large scale comparison of random forest and a feature tokenizer transformer, with explainability and fairness analysis, at an Italian university; and a four step logistic regression study of dropout intention among student teachers in Ethiopia We compare their dropout definitions, prediction timing, data, imbalance handling, algorithms, metrics and explainability practices. The synthesis shows that predictivepower rises sharply once post enrolment academic or engagement data become available, that tree based and ensemblemodels remain competitive on tabular institutional data, that recall oriented evaluation is often reported without calibratedprecision, and that results are rarely comparable across studies. We propose a stage aware, human in the loop framework and aminimum reporting checklist to improve the comparability, fairness and practical usefulness of future dropout prediction research.