Systematic review on the use of machine learning to detect school learning difficulties

Abstract

This systematic literature review (SLR) analyzes the impact of machine learning on the early detection of learning difficulties in school settings. Using the PICO methodology and the PRISMA protocol, four key questions were articulated regarding types of difficulties, applied algorithms, comparison with traditional methods, and intervention improvements. A search of the Scopus database identified 306 documents, from which 36 relevant studies were selected after applying rigorous inclusion and exclusion criteria. The findings show that algorithms such as Random Forest, SVM, deep neural networks, and ensemble models allow the identification of complex patterns in academic, behavioral, and neuropsychological data, far exceeding the accuracy and agility of conventional methods. The main applications include the detection of dyslexia, dysgraphia, dyscalculia, and ADHD, as well as the prediction of academic risk and pedagogical personalization. Furthermore, it is observed that sensory and adaptive artificial intelligence tools strengthen educational inclusion for students with cognitive disabilities. However, significant challenges remain, such as the need for high-quality data, limited validation in real-life school settings, and the poor interpretability of some complex models. In conclusion, the use of machine learning represents an effective solution for improving the early detection of learning difficulties, overcoming the limitations of traditional approaches and opening up new opportunities for more accurate, timely, and inclusive educational interventions.