Classification Mental Workload Levels from EEG Signals with 1D Convolutional Neural Network

Baydemir, Recep; Latifoğlu, Fatma; Orhanbulucu, Fırat · 2022 · Crossref

DOI: 10.56038/ejrnd.v2i4.193

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 accurately classifying mental workload (MWL) levels from electroencephalogram (EEG) signals, a critical task for brain-computer interfaces and driving system evaluations. While previous methods relied on manual feature extraction and traditional machine learning classifiers like Support Vector Machines (SVM), which achieved success rates ranging from 79.8% to 95.39%, this research investigates the efficacy of a one-dimensional convolutional neural network (1D-CNN) for automatic feature extraction and classification. The primary motivation is to leverage deep learning’s ability to process raw data directly, potentially surpassing the performance of existing SVM-based approaches on the same dataset. The experimental design utilized the STEW dataset provided by Lim et al., comprising EEG recordings from 48 male participants performing the SIMKAP multitasking test. Signals were recorded from 14 channels at 128 Hz. Participants were categorized into low (L) and high (H) workload groups based on post-experiment questionnaires. The methodology involved two distinct classification conditions. In the first condition, raw EEG signals were preprocessed with a 0.4–30 Hz band-pass filter and segmented into two equal parts to augment data volume. In the second condition, these signals were further decomposed into six subbands using Empirical Mode Decomposition (EMD) to increase data diversity. Both conditions employed a 1D-CNN model trained with the Adam optimizer, ReLU activation functions, and a mini-batch size of 64. Performance was evaluated using 5-fold cross-validation and a hold-out technique, measuring accuracy (Acc), sensitivity (Sens), and specificity (Spe). The results demonstrated that the 1D-CNN model achieved superior performance in the first condition (preprocessed raw signals) compared to the EMD-decomposed signals in the second condition. The highest accuracy recorded was 98.4%, with a sensitivity of 97.62% and specificity of 98.94%. In contrast, the EMD-based approach yielded a lower average accuracy of 92.42%. Although EMD increased the total number of data points from 1,260 to 7,560, the decomposition process appeared to reduce classification success rates relative to the raw signal approach. When compared to prior studies using the same dataset, which reported accuracies between 69% and 95.9%, the proposed 1D-CNN model outperformed existing methods. The significance of this work lies in demonstrating that 1D-CNNs can effectively classify binary MWL levels without manual feature engineering, achieving near-perfect accuracy on the STEW dataset. The findings suggest that while signal decomposition techniques like EMD can augment data volume, they may introduce complexity that hinders classification performance in this specific context. This study contributes to the field by validating deep learning architectures for EEG-based workload estimation and highlighting the trade-offs between data augmentation strategies and model accuracy in cognitive state classification.

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