The rapid development of e-learning platforms has increased demand for adaptive learning models that enhance engagement and personalization. Adaptive e-learning focuses on understanding the individual’s learning styles, abilities, and adjusting accordingly. This paper proposes a modular hybrid adaptive approach for engagement prediction in e-learning systems, integrating ensemble learning paradigms, bagging, boosting, and stacking to enhance model performance and accuracy. The architecture employs Random Forest (RF), and CatBoost (CB), as base learners, leveraging bagging to reduce variance and boosting to minimise bias, respectively. These models are then combined in a novel stacking ensemble framework, where LightGBM acts as the meta-learner, aggregating the outputs of RF and CB to further improve generalization through a diverse learning mechanism. The Open University Learning Analytics dataset, which includes learner performance, contentment, and course interaction behaviors, was used in the study. Experimental results demonstrate that the proposed hybrid adaptive ensemble approach outperforms the individual models. RF achieved 93.8% accuracy, CatBoost 97.25%, and the stacking model 99.8% training accuracy and testing accuracy of 97.86%, with a minimal generalization gap of 2.2%, depicting strong learning capacity with low overfitting and improved generalization. These findings underscore the effectiveness of ensembled stack-based adaptability in advancing personalized e-learning platforms.