Using Deep Learning in Student Performance Prediction: A Systematic Review

This article examines how Deep Learning (DL) techniques can predict student achievement (SA), following the methodology of Kitchenham and Charters. The study focuses on identifying the most widely used models, assessing their accuracy, and determining key variables for SA prediction. From 1,096 studies reviewed, 27 were selected that met the inclusion criteria, such as focusing specifically on DL and providing details about the models and variables used. The most common models were Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Deep Neural Networks (DNN), with an accuracy of over 90%. It is highlighted that, although these models are effective, they face challenges such as the small amount of data and representativeness, the analysis of "black box" models, and the generalization of results to diverse educational contexts. In conclusion, DL has great potential to improve SA prediction and the early identification of at-risk students but requires overcoming important challenges to maximize its impact on education.