Explainable learning analytics to predict college dropout rates in hybrid settings

Abstract

University dropout in hybrid environments is a multidimensional problem whose prediction requires integrating academic trajectories with students’ socioeconomic, psychological, institutional, and digital conditions. The objective was to analyze and systematize the evidence needed to guide the development and evaluation of explainable learning analytics applied to Latin American universities. A qualitative, documentary, and cross-sectional study was conducted through a structured search in five academic databases, the selection of recent publications, and deductive content analysis focused on risk factors, algorithms, explainability, calibration, and equity. The results showed that dropout was associated with the accumulation of risks rather than with isolated indicators of digital participation; likewise, ensemble models achieved aggregated accuracy levels of approximately 86–88%, although only about 24% of the studies incorporated explainability techniques. The evidence also revealed limited temporal, institutional, and intersectional validation, as data from a single institution predominated. These findings indicate that predictive accuracy does not guarantee pedagogically relevant interpretations or equitable decisions. It was concluded that the usefulness of these systems depends on integrating performance, explainability, calibration, transferability, human oversight, and institutional responsibility.