Attention-guided deep learning-machine learning and statistical feature fusion for interpretable mental workload...

Accurate assessment of mental workload (MWL) from electroencephalography (EEG) signals is crucial for real-time cognitive monitoring in safety-critical domains such as aviation and human-computer interaction. Although various computational approaches have been proposed, those mostly suffer from limited robustness, interpretability, or fail to fully exploit both temporal and non-linear neural dynamics. This article introduces a novel hybrid deep learning and XGBoost stacking ensemble framework for reliable and interpretable MWL classification from EEG. The proposed pipeline systematically includes preprocessing of raw EEGs, followed by comprehensive feature extraction (time-domain, frequency-domain, wavelet-based, entropy, and fractal dimension features), and subsequent discriminative feature selection phase using ANOVA F-values, yielding a compact set of 200 highly informative features. The proposed architecture consists of dual processing branches: a CNN-BiLSTM-Attention based deep learning branch for automatic learning of spatiotemporal dynamics, and an XGBoost branch for robust classification from engineered features. Predictions from both branches are integrated using a logistic regression stacking ensemble, maximizing complementary strengths and improving generalization. Experiments are conducted on the STEW (simultaneous workload) and EEGMAT (mental arithmetic task) dataset. Proposed model yields 96.87% and 99.40% of classification accuracy by outperforming 16 and 7 previously published state-of-the-art techniques on STEW and EEGMAT dataset respectively. Attention heatmaps and SHAP value analysis provide intuitive visual explanations and interpretability of the model's decision making, while systematic ablation studies validate the contribution of each architectural module. This work demonstrates that a carefully engineered stacking ensemble, informed by both deep learning and classical machine learning, capable of delivering not only improved performance but also enhanced interpretability for EEG-based MWL assessment in real-world applications.