Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Predicting Workload in Virtual Flight Simulations using EEG Features (Including Post-hoc Analysis in Appendix)

Bas Verkennis, Evy van Weelden, Francesca L. Marogna, Maryam Alimardani, Travis J. Wiltshire, Max M. Louwerse

TL;DR

Problem: Real-time workload estimation for pilots during flight training using EEG BCIs. Approach: compare spectral power features with fronto-parietal PLV connectivity in VR and Desktop flight simulations, using a stacked classifier and NASA-TLX labels. Findings: integrating spectral and connectivity features yields higher accuracy (78% vs 50%), with alpha/beta PLV connectivity and parietal-frontal links driving performance; post-hoc analyses mitigate confounds and mixed-effects residuals improve label robustness. Significance: supports PLV-based workload prediction as a route to adaptive VR training systems in safety-critical aviation contexts, and points to future work with larger, homogeneous pilot samples and additional connectivity metrics.

Abstract

Effective cognitive workload management has a major impact on the safety and performance of pilots. Integrating brain-computer interfaces (BCIs) presents an opportunity for real-time workload assessment. Leveraging cognitive workload data from high-fidelity virtual reality (VR) flight simulations allows for dynamic adjustments to training scenarios. While prior studies have predominantly concentrated on EEG spectral power for workload prediction, delving into intra-brain connectivity may yield deeper insights. This study assessed the predictive value of EEG spectral and connectivity features in distinguishing high vs. low workload periods during simulated flight in VR and Desktop conditions. Using an ensemble approach, a stacked classifier was trained to predict workload from the EEG signals of 52 participants. Results showed that the mean accuracy of the model incorporating both spectral and connectivity features improved by 28% compared to the model that solely relied on spectral features. Further research on other connectivity metrics and deep learning models in a large sample of pilots is essential to validate the potential of a real-time workload-prediction BCI. This could contribute to the development of an adaptive training system for safety-critical operational environments.

Predicting Workload in Virtual Flight Simulations using EEG Features (Including Post-hoc Analysis in Appendix)

TL;DR

Problem: Real-time workload estimation for pilots during flight training using EEG BCIs. Approach: compare spectral power features with fronto-parietal PLV connectivity in VR and Desktop flight simulations, using a stacked classifier and NASA-TLX labels. Findings: integrating spectral and connectivity features yields higher accuracy (78% vs 50%), with alpha/beta PLV connectivity and parietal-frontal links driving performance; post-hoc analyses mitigate confounds and mixed-effects residuals improve label robustness. Significance: supports PLV-based workload prediction as a route to adaptive VR training systems in safety-critical aviation contexts, and points to future work with larger, homogeneous pilot samples and additional connectivity metrics.

Abstract

Effective cognitive workload management has a major impact on the safety and performance of pilots. Integrating brain-computer interfaces (BCIs) presents an opportunity for real-time workload assessment. Leveraging cognitive workload data from high-fidelity virtual reality (VR) flight simulations allows for dynamic adjustments to training scenarios. While prior studies have predominantly concentrated on EEG spectral power for workload prediction, delving into intra-brain connectivity may yield deeper insights. This study assessed the predictive value of EEG spectral and connectivity features in distinguishing high vs. low workload periods during simulated flight in VR and Desktop conditions. Using an ensemble approach, a stacked classifier was trained to predict workload from the EEG signals of 52 participants. Results showed that the mean accuracy of the model incorporating both spectral and connectivity features improved by 28% compared to the model that solely relied on spectral features. Further research on other connectivity metrics and deep learning models in a large sample of pilots is essential to validate the potential of a real-time workload-prediction BCI. This could contribute to the development of an adaptive training system for safety-critical operational environments.

Paper Structure

This paper contains 20 sections, 2 equations, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Dataset Description
  4. Binary Classification of Workload using NASA-TLX
  5. Data Pre-processing
  6. Feature Extraction
  7. Spectral Power
  8. Phase Locking Value (PLV)
  9. Feature Selection
  10. Classification Models
  11. Model Evaluation
  12. Results
  13. Discussion
  14. Limitations
  15. Future Work
  16. ...and 5 more sections

Figures (1)

  • Figure 1: Overview of the BCI classification pipeline. EEG = electroencephalogram; PLV = Phase Locking Value; RFE = Recursive Feature Elimination; Stacked Classifier = ensemble model combining Random Forest, Logistic Regression, and SVM as base models, with SVM used as the meta-model; NASA-TLX = NASA Task Load Index.