A systematic review of deep learning techniques for learners’ engagement prediction in electronic learning systems

Abstract

Electronic learning enables learners to acquire knowledge outside the confines of a classroom. However, engagement is of significant concern, as the system does not facilitate one-on-one monitoring. To propose innovative systems capable of forecasting learners’ engagement, it is essential to conduct an in-depth investigation to ascertain the extent to which researchers have explored this domain. This research presents a comprehensive systematic review on e-learners’ engagement prediction, focusing on sources of data acquisition, employed features, deep learning techniques utilized, and the performances achieved by various techniques. The search across scholarly databases including Springer, Science Direct and Web of Science returned 301 articles. The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) methodology was used to screen the articles focusing on studies published between 2019 and 2025. However, after an extensive examination of the articles only 42 satisfied the inclusion criteria which were included in the final review. Research questions that examined the sources of data used for engagement prediction, feature sets employed, the deep learning algorithms leveraged, the relevant challenges, and potential future research paths in this field guided the research. The findings from the investigation revealed that most researchers utilize Dataset for Affective States in E-learning Environments (DAiSEE) datasets in schooling-related models, and learners’ emotions are frequently employed. Moreover, Convolutional Neural Network architectures are predominantly leveraged for the development of engagement predictive models. Recurrent Neural Network-based architectures are combined with Convolutional Neural Network architectures to create hybrid models. The study observed the highest accuracy of 99%, achieved by a Convolutional Neural Network model. Key challenges were identified from the investigation including heavy reliance on single-modal data, especially learners’ facial emotions, and narrow model generalization. However, key future research directions were identified including the development of multi-modal engagement datasets and models that are computationally efficient for real-time deployment.