Learning Analytics-Based Early Warning System for Predicting Student Dropout Risk in Asynchronous Online Learning

Abstract

Student dropout remains a major challenge in asynchronous online learning due to limited direct interaction and the difficulty of identifying disengaged learners at an early stage. This study aimed to develop and evaluate a Learning Analytics-Based Early Warning System for predicting student dropout risk by integrating behavioral learning analytics, machine learning, and explainable artificial intelligence. A quantitative Design Science Research approach was employed to design, implement, and evaluate the proposed system. Behavioral data were collected from a Learning Management System, including login frequency, learning inactivity, learning resource utilization, assignment submission, discussion participation, and other engagement indicators. Multiple machine learning algorithms, including Random Forest, XGBoost, LightGBM, LSTM, and BLSTM, were evaluated using 10-fold cross-validation based on Accuracy, Precision, Recall, F1-score, and ROC-AUC. The findings indicate that behavioral learning data effectively distinguish students with different levels of dropout risk. XGBoost achieved the best overall predictive performance, while SHapley Additive exPlanations enhanced model transparency by identifying the behavioral factors contributing to prediction outcomes. Assignment submission regularity, login frequency, learning inactivity, and discussion participation emerged as the most influential predictors. The developed Early Warning System integrates automated prediction, behavioral visualization, explainable analytics, and personalized intervention recommendations within a single decision-support platform. The proposed framework provides a practical and scalable solution for improving early identification of at-risk students, supporting timely academic intervention, and strengthening student retention in asynchronous online learning environments.