Discrimination of Mental Workload Levels From Multi-Channel fNIRS Using Deep Leaning-Based Approaches

Ho, Thi Kieu Khanh; Gwak, Jeonghwan; Park, Chang Min; Song, Jong-In · 2019 · Crossref

DOI: 10.1109/access.2019.2900127

archive: archived pipeline: cataloged

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

Summary

**Summary** This study addresses the challenge of distinguishing mental workload levels (rest vs. task) using functional near-infrared spectroscopy (fNIRS), a non-invasive neuroimaging technique that measures hemodynamic changes in the prefrontal cortex. While fNIRS is portable and suitable for continuous monitoring, its signals are often non-stationary and suffer from low signal-to-noise ratios, making traditional feature extraction difficult. The research aims to evaluate the efficacy of deep learning approaches compared to conventional machine learning methods for classifying these mental states, specifically focusing on oxy-hemoglobin (HbO), deoxy-hemoglobin (HbR), and their difference (HbT). The experimental design involved 16 healthy volunteers (8 male, 8 female, aged 18–30) performing Stroop task experiments (STEs) to induce varying levels of cognitive stress. fNIRS data was recorded using a Shimadzu FOIRE-3000 system with seven channels over the prefrontal cortex. The protocol included congruent and incongruent word stimuli, with incongruent words comprising 83.69% of trials to trigger higher workload. Data preprocessing involved converting raw signals to hemoglobin concentrations, applying a 0.3–3 Hz bandpass filter, and segmenting data into non-overlapping windows. Principal Component Analysis (PCA) was applied to reduce dimensionality, retaining 95% of the variance. Four classifiers were tested: Support Vector Machine (SVM), Adaptive Boosting (AdaBoost), Deep Belief Network (DBN), and Convolutional Neural Network (CNN). The results demonstrated that deep learning models outperformed traditional machine learning algorithms in classification accuracy. The DBN achieved the highest accuracy of 84.26% ± 2.58%, followed by the CNN at 72.77% ± 1.92%. In contrast, the conventional AdaBoost and SVM methods yielded lower accuracies of 71.13% ± 2.96% and 64.74% ± 1.57%, respectively. The study confirmed that apparent changes in HbO and HbR concentrations between rest and task periods exist across participants, validating fNIRS as a viable modality for workload assessment. The significance of this work lies in its demonstration that deep learning architectures, particularly DBNs, can effectively handle the noise and non-stationarity inherent in fNIRS signals without relying on hand-crafted features. This suggests that fNIRS-based Brain-Computer Interfaces (BCIs) can be made more robust and practical for real-world applications, such as monitoring driver fatigue or cognitive load in everyday tasks, by leveraging automated feature extraction through deep neural networks.

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

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