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

Area of Science:

• Neuroscience
• Machine Learning
• Human-Computer Interaction

Background:

• Accurate mental workload (MWL) assessment using electroencephalography (EEG) is vital for real-time cognitive monitoring in critical domains.
• Existing computational methods often lack robustness, interpretability, or fail to capture complex neural dynamics.

Purpose of the Study:

• To introduce a novel hybrid deep learning and XGBoost stacking ensemble for reliable and interpretable MWL classification from EEG.
• To enhance the accuracy and generalization of MWL assessment by integrating complementary strengths of different machine learning approaches.

Main Methods:

• A hybrid framework combining a CNN-BiLSTM-Attention deep learning branch and an XGBoost branch, integrated via a logistic regression stacking ensemble.
• Comprehensive EEG preprocessing, feature extraction (time, frequency, wavelet, entropy, fractal dimensions), and feature selection (ANOVA F-values) yielding 200 features.
• Validation on STEW and EEGMAT datasets, employing attention heatmaps and SHAP values for interpretability analysis.

Main Results:

• Achieved 96.87% accuracy on the STEW dataset and 99.40% on the EEGMAT dataset.
• Outperformed 16 and 7 previously published state-of-the-art techniques on the respective datasets.
• Demonstrated enhanced interpretability through attention heatmaps and SHAP value analysis.

Conclusions:

• The proposed hybrid deep learning and XGBoost stacking ensemble framework offers superior performance and interpretability for EEG-based MWL assessment.
• This approach effectively leverages both deep learning for spatiotemporal dynamics and classical machine learning for engineered features, improving real-world applicability.