Mental workload (MWL) classification using electroencephalogram (EEG) signals is crucial for cognitive neuroscience and is also a challenging research area in brain-computer interface (BCI).
Since the EEG signals fluctuate a lot across sessions and individuals, there is a need for a robust classification model that generalizes well for real-world applications.
Furthermore, Shapley Additive Explanations (SHAP) analysis was performed to identify the most influential EEG channels for model prediction.
Finally, the SHAP analysis revealed that the most informative EEG channels were located over the parieto-occipital and temporal regions, which is consistent with established neurophysiological evidence related to MWL processing.
The proposed framework improves both performance and explainability in EEG-based MWL classification, representing a systematic integration of SMOTE and SHAP analysis.
Mental workload (MWL) classification using electroencephalogram (EEG) signals is crucial for cognitive neuroscience and is also a challenging research area in brain-computer interface (BCI). Since the EEG signals fluctuate a lot across sessions and individuals, there is a need for a robust classification model that generalizes well for real-world applications. In this work, we used the publicly available dataset “An EEG dataset for cross-session mental workload estimation: passive BCI competition of the Neuroergonomics Conference 2021”, and the standard EEGNet model to classify the MWL into three classes (Low, Med, and High). To improve the performance of the model, a synthetic minority oversampling technique (SMOTE) was used by creating synthetic EEG samples, and key hyperparameters (F 1 , F 2 , and D) of EEGNet were systematically varied to identify the optimal configuration. Furthermore, Shapley Additive Explanations (SHAP) analysis was performed to identify the most influential EEG channels for model prediction. The proposed approach achieves the highest accuracy of 80.5% and 82.7% without and with SMOTE, respectively. The comparative analysis showed that applying SMOTE resulted in an average performance improvement of approximately 3%. A Wilcoxon signed-rank test confirmed that this improvement was statistically significant (p < 0.05). Finally, the SHAP analysis revealed that the most informative EEG channels were located over the parieto-occipital and temporal regions, which is consistent with established neurophysiological evidence related to MWL processing. The proposed framework improves both performance and explainability in EEG-based MWL classification, representing a systematic integration of SMOTE and SHAP analysis.
