Hybrid deep learning for mental workload classification using EEG with enhanced preprocessing and interpretability
DOI: 10.1371/journal.pone.0352882
archive: archived pipeline: cataloged
Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)
Summary
This paper addresses the challenges of generalizability, noise robustness, and interpretability in electroencephalography (EEG)-based mental workload (MWL) classification, a critical task in safety-sensitive fields like healthcare and aviation. While deep learning has improved upon traditional machine learning methods, existing hybrid architectures often rely on static representations or weaker evaluation protocols, leaving a gap in robust, subject-independent frameworks that provide neurophysiological insight. To address this, the authors propose a unified hybrid deep learning framework that integrates a Variational Autoencoder (VAE), a Convolutional Block Attention Module (CBAM), and a Bidirectional Long Short-Term Memory (BLSTM) network. The study utilized the Simultaneous Task EEG Workload (STEW) dataset, comprising recordings from 48 participants performing the SIMKAP multitasking test. EEG signals were acquired using a 14-electrode Emotiv EPOC headset and categorized into four workload levels: Baseline, Low, Medium, and High. The methodology involved preprocessing raw EEG data through bandpass filtering (1–40 Hz) and artifact removal, followed by decomposition into five frequency bands (delta, theta, alpha, beta, gamma). These bands were transformed into topographical videos using Morlet wavelet analysis. The VAE processed these videos to extract noise-reduced latent features, which were then refined by the CBAM to focus on informative spatial-channel patterns. Finally, the BLSTM captured temporal dependencies to classify the workload states. The model was evaluated using strict leave-one-subject-out (LOSO) cross-validation to ensure cross-subject generalization. The proposed VAE-CBAM-BLSTM framework achieved an average accuracy of 83.9% (±3.1%) and a macro-F1 score of 0.83 (±0.04) across subjects, outperforming baseline models. Ablation studies confirmed that each architectural component—VAE, CBAM, and BLSTM—contributed significantly to performance improvements. Sensitivity analyses identified a 10-second window length as optimal for balancing temporal context and label specificity. Furthermore, Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations revealed that the model’s predictions were driven by activity in frontal-parietal brain regions and specific frequency bands, providing neurophysiological interpretability. The VAE demonstrated stable training with clear clustering in the latent space, indicating effective denoising and class separability. The significance of this work lies in its provision of a robust, interpretable, and generalizable framework for EEG-based MWL assessment. By integrating denoising, attention-based feature refinement, and temporal modeling under strict LOSO validation, the study offers a practical solution for real-world applications where cross-subject variability and signal noise are major hurdles. The inclusion of interpretability analysis bridges the gap between black-box deep learning models and neurophysiological understanding, suggesting that the identified brain regions align with known workload dynamics. Future research directions include exploring adaptive windowing strategies, multimodal data integration, and benchmarking against transformer-based and graph neural network architectures to further enhance generalizability.
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.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | Crossref | — | — | 1 | 2026-08-09 |
| archive | success | canonical_url | — | — | 1 | 2026-08-09 |
| extract | success | cached | — | — | 4 | 2026-08-23 |
| clean | success | clean | — | — | 1 | 2026-08-09 |
| chunk | success | chunk | — | — | 1 | 2026-08-09 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-08-09 |
| promote | success | — | — | — | 1 | 2026-08-09 |
| summarize | success | llm | qwen3.8-27b-gittensor | summ-v5 | 3 | 2026-08-23 |
| tag | success | vector_similarity | — | — | 11 | 2026-08-11 |
| verify | success | — | — | — | 2 | 2026-08-09 |
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.
Information type
What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).
- Empirical Findings: physiological data