Explainable Artificial Intelligence for Cognitive Workload Assessment: Methods, Evidence, and Human-Centered Evaluation
Cognitive workload models now draw on eye movements, brain activity, cardiovascular variation, electrodermal responses, and patterns of interaction. Their accuracy can be impressive, yet accuracy alone says little about whether the model has captured workload, memorized a participant or task, or produced an explanation that a person can safely use. This critical narrative review examines that problem across cognitive science, neuroergonomics, physiological computing, and explainable artificial intelligence (XAI). Particular attention is given to construct definition, signal-specific confounding, participant-wise validation, and the gap between plausible and faithful explanations. The review develops TRACE-WL: Theoretical alignment, Reliable validation, Attribution fidelity, Cognitive usability, and Ecological robustness. These five requirements link the psychological claim to the validation design and, ultimately, to the conditions of use. The literature indicates that predictive performance and explanation quality must be tested separately. It also cautions against treating feature importance as evidence of a cognitive mechanism without convergent measurements or intervention. TRACE-WL offers a structured way to judge workload systems intended for education, healthcare, transportation, and other human-machine settings.
Authors
- Napat Sukthong (ORCID: https://orcid.org/0009-0007-5405-6068)
Publication Details
- Journal
- Preprints.org
- Published
- 2026-11-18
- DOI
- https://doi.org/10.20944/preprints202611.1808.v1
- Primary Topic
- Human-Automation Interaction and Safety
- Type
- preprint