SHAP analysis of an improved EEG-based mental workload classification framework: utilizing data augmentation and explainable AI

Chaturvedi, Sushil; Ahirwal, Mitul Kumar · 2026 · Crossref

DOI: 10.1038/s41598-026-52330-z

archive: archived pipeline: cataloged

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of classifying mental workload (MWL) from electroencephalogram (EEG) signals, a critical task for brain-computer interfaces and cognitive neuroscience. The primary motivation is the poor generalizability of models trained on single-session data due to the non-stationary nature of EEG signals across different days and individuals. To improve robustness and interpretability, the authors propose a framework that integrates the standard EEGNet architecture with Synthetic Minority Oversampling Technique (SMOTE) for data augmentation and Shapley Additive Explanations (SHAP) for explainable artificial intelligence (XAI). The methodology utilizes the publicly available "An EEG dataset for cross-session mental workload estimation" from the Neuroergonomics Conference 2021 passive BCI competition. This dataset comprises EEG recordings from 15 participants performing the Multi-Attribute Task Battery–II (MATB-II) task across three sessions, each with low, medium, and high difficulty levels. The EEG data, recorded via 62 electrodes at 500 Hz, was preprocessed using EEGLAB, including bandpass filtering (1–40 Hz) and downsampling to 250 Hz. The classification model is EEGNet, a lightweight deep learning architecture. The researchers systematically varied key hyperparameters (F1, F2, and D) across 10 configurations, testing each with and without SMOTE. SMOTE was applied at the epoch level to the training set to address class imbalance, while SHAP analysis was employed to identify the most influential EEG channels for model predictions. The results demonstrate that SMOTE consistently improves classification performance. The highest accuracy achieved was 82.7% with SMOTE, compared to 80.5% without it, representing an average improvement of approximately 3%. A Wilcoxon signed-rank test confirmed that this improvement was statistically significant (p < 0.05) for 8 out of 10 hyperparameter configurations. The optimal configuration was identified as F1 = 6, F2 = 128, and D = 4, which yielded approximately 83% accuracy with SMOTE. Furthermore, SHAP analysis revealed that the most informative EEG channels were located over the parieto-occipital and temporal regions. This finding aligns with established neurophysiological evidence regarding MWL processing, thereby enhancing the model's interpretability. The significance of this work lies in its systematic integration of data augmentation and explainable AI within a cross-session EEG framework. By addressing data scarcity through SMOTE and providing neurophysiological insights via SHAP, the study offers a more robust and transparent approach to MWL classification. This framework not only improves accuracy and stability in real-world, multi-session applications but also bridges the gap between black-box deep learning models and neuroscientific understanding, facilitating the development of reliable brain-computer interfaces for monitoring cognitive states in sectors such as aviation, healthcare, and driving.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.8-27b-gittensor on 2026-08-23; verification: pending re-verification.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• neuro workload indices