Distance-learning platforms increasingly seek objective EEG-based markers of lecture comprehension, yet robust, imbalance-aware models on real datasets remain scarce. Using the Kaggle “EEG data / Distance learning environment” corpus (8 students, 14 Emotiv Epoc X channels, 84 tabular features/segment, ≈21% “not understood”), we propose IASTE, an Imbalance-Aware Stacked Tree Ensemble combining gradient boosting, RUSBoost, and bagged trees via leakage-free out-of-fold stacking and a logistic meta-learner. Under a stratified 70/15/15 train–validation–test split, with hyperparameters selected solely by minority-class F1 on validation, IASTE attains 97.3% test accuracy, macro-F1 = 0.959, and F1₀ (“not understood”) = 0.936. This improves macro-F1 by ≈1–2 percentage points and F1₀ by up to ≈1.5 points over strong baselines including Random Forest, RUSBoost, SVM, LSTM, BiLSTM, and BiLSTM+attention. Subject-wise analysis shows per-student accuracies in the range 0.968–0.980, versus ≈0.956–0.964 for tree and deep baselines, indicating genuine cross-subject generalization. Ablations confirm that removing stacking, imbalance handling, or F1₀-based selection systematically degrades minority-class F1 and macro-F1, while performance remains stable across the explored tree-depth and NumLearningCycles grid. By enabling objective, data-driven monitoring of lecture comprehension in distance-learning environments, our approach supports more inclusive and effective digital education.

EEG, Distance learning, Lecture understanding, Class imbalance, Stacked ensembles, Subject-wise generalization.