The rapid expansion of e-learning has resulted in a surge in educational data volume, presenting challenges in manually uncovering valuable information. Concurrently, advancements in educational data mining offer robust technical support for forecasting student performance based on their engagement behaviors. In this study, we initially investigate the efficacy of machine learning algorithms for forecasting student performance, followed by an exploration of the application of deep learning techniques in this domain. This research proposes the utilization of a Large-scale Feature-Derived Bidirectional Long Short-Term Memory Neural Network (LFD-BiLSTM) model for predicting student performance. We analyze students' learning behaviors using clickstream data extracted from virtual learning environments and investigate the primary factors influencing academic achievement. The results indicate that the model achieves a high accuracy of 92.57%, a recall rate of 93.90%, and an F1 score of 92.96%. Furthermore, the experimental findings underscore the significant influence of students' learning behavior, educational level, and the Index of Multiple Deprivation on their academic performance, providing valuable insights for educators, policymakers, and researchers to tailor individualized programs.