Unsupervised classification of operator workload from brain signals.

Area of Science:

• Neuroscience
• Human-Computer Interaction
• Cognitive Engineering

Background:

• Operator workload assessment is crucial for real-world environments.
• Brain-signal based predictors offer potential for objective workload measurement.
• Developing methods that require minimal or no labeled data is essential for practical applications.

Purpose of the Study:

• To classify operator workload using brain signals.
• To explore predictors with varying levels of required label information, including unsupervised approaches.
• To investigate the fusion of brain signals and peripheral physiological measures (PPMs) for enhanced classification.

Main Methods:

• Subjects performed a visually and motorically demanding task with alternating difficulty.
• Compared classical electroencephalography (EEG) sensor-space methods with spatial filtering techniques: Common Spatial Patterns (CSP), Source Power (SPoC), and canonical SPoC (cSPoC).
• Investigated the impact of combining brain signals with PPMs on classification performance.

Main Results:

• Spatial filtering methods achieved high classification accuracies: CSP (94%), SPoC (92%), and cSPoC (82%).
• These methods outperformed non-spatial filtering approaches and identified physiologically plausible components.
• Unsupervised cSPoC performance significantly improved when augmented with PPM features.

Conclusions:

• Workload states can be successfully differentiated using brain signals, even with limited or no labeled data.
• Spatial filtering techniques, particularly CSP and SPoC, are effective for workload classification.
• Fusing brain signals with PPMs enhances classification accuracy, especially for unsupervised methods, paving the way for real-world applications.