A machine learning-based adaptive platform is associated with cognitive improvements in Peruvian secondary school students: a quasi-experimental study

Abstract

This study determined the impact of a personalized educational platform based on machine learning (ML) on the cognitive learning of secondary education students at a public institution in Piura, Peru. The research followed an applied, quantitative approach with a quasi-experimental design. A sample of 56 fourth-grade students was distributed into a control group ( n  = 26, traditional teaching) and an experimental group ( n  = 30, platform-assisted teaching), using two pre-existing classrooms randomly assigned to each condition. Cognitive learning was assessed across four dimensions (attention, memory, reasoning, and comprehension) with pretest and posttest exams administered at the start and end of the same two-hour session per topic, over three consecutive weeks. The platform combined an unsupervised clustering model for cognitive profiling with a supervised classifier that predicted, in real time, the optimal content level for each student. Analyses were based on improvement scores (posttest - pretest); intergroup comparisons used the Mann–Whitney U test with rank-biserial effect sizes, and the relationship between platform metrics and the posttest score was examined through Spearman correlations. Results showed significant differences favoring the experimental group across the three topics ( p  < .001), with increasing effect sizes (r = 0.527, 0.610, 0.738), confirmed through a sensitivity analysis that controlled for the pretest. The experimental group achieved an average improvement of 8.29 points out of 20, versus 4.90 for the control group. Requested adaptations showed a moderate association on average with the posttest score ( ρ  = 0.485, p  < .001), strengthening over the course of the intervention (up to ρ  = 0.588), while the predictive weight of the initial micro-diagnostic declined progressively. The platform was associated with a significant, progressive, and dimension-differentiated improvement, although the single-classroom-per-condition design and the within-session measurement warrant interpretive caution. This work contributes to inclusive education in resource-limited contexts and aligns with Sustainable Development Goals 4 and 10.