Massive Open Online Courses (MOOCs) have gained popularity as an accessible form of education, attracting a diverse and widespread student base. Despite their potential, MOOCs face a significant challenge: high dropout rates, which undermine their effectiveness and impact. The increasing interest in addressing this problem led to numerous studies developing new models to predict dropouts early and automatically, many of which use Machine Learning (ML) approaches. This research performs a quantitative synthesis of the performance of ML techniques for early dropout prediction in MOOCs. Following PRISMA guidelines, we perform a systematic review and meta-analysis. To analyze the overall performance, we use a random-effects model for a meta-analysis of proportions, analyzing two metrics: sensitivity and specificity. We have also studied the relationship between some of the studies’ characteristics and the performance obtained by means of subgroup analysis. The results indicate that ML systems are capable of accurately detecting a significant percentage of potential dropouts. However, the performance of these systems varies depending on the dataset and the definition of dropout used in each study. Despite the promising findings, the high heterogeneity observed across studies suggests that these results should be interpreted with caution.
Tenorio-Berrio et al. (2026) studied this question.