Time-Series Foundation Models for Cognitive Workload Classification using Eye-Tracking Data
- 1University of Central Florida
Abstract
Cognitive workload (CWL) assessment is relevant to a range of applications, such as monitoring driver fatigue and pilot attention, surgeon workload during complex procedures, and astronaut cognitive fatigue during long-duration missions. Eye-tracking datasets are generally small, which hinders the generalizability of the AI models. Time-series foundation models (TSFMs) have shown promise in mitigating this limitation as they are pretrained on large corpora and can be effectively fine-tuned with limited task-specific data. Although cognitive workload measurement has been explored in controlled settings, the ability of TSFMs to generalize to unseen individuals for CWL classification from eye-tracking data has not been studied. In this paper, two TSFMs, MOMENT and Moirai, were evaluated for CWL classification and compared against CNN, FFN, and LSTM baselines on two publicly available eye-tracking datasets. We used subject-level five-fold cross-validation in which data from each test subject were held out during training. We report accuracy, AUC, and F1-score with 95% confidence intervals. On the class-balanced EM-COGLOAD dataset, the pretrained TSFMs generalized markedly better to unseen subjects, with Moirai reaching 0.918 AUC and MOMENT 0.882 AUC, while the task-specific baselines remained near 0.70 AUC. On the class-imbalanced COLET dataset, even though the performance of all the models decreased, MOMENT was the most robust (0.670 AUC). Both TSFMs were also the most stable across held-out subjects, indicating that pretrained representations generalize more consistently across individuals than task-specific models in data-scarce eye-tracking settings.
Related
- Benchmarking foundation models for resting-state functional magnetic resonance imaging time series in cognitive aging research
- Cognitive workload across subjective and electrodermal measures in relation to previous-night sleep
- Tracking acquisition of cognitive maps in early- and late-onset blindness
- An adolescent neuroimaging database combining movie-watching, eye-tracking and cognitive tasks
- Repeated within-session rule switching in the marmoset: a paradigm for tracking the dynamics of cognitive flexibility
BibTeX
@misc{haag2026timeseries,
title = {{Time-Series Foundation Models for Cognitive Workload Classification using Eye-Tracking Data}},
author = {J. Haag and J. G. Gonzalez Nunez and B. Kirkwood and G. Legault and L. Brattain},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.13.751282},
url = {https://www.biorxiv.org/content/10.64898/2026.09.13.751282v1},
note = {CC BY-ND},
}
RIS
TY - UNPB
TI - Time-Series Foundation Models for Cognitive Workload Classification using Eye-Tracking Data
AU - J. Haag
AU - J. G. Gonzalez Nunez
AU - B. Kirkwood
AU - G. Legault
AU - L. Brattain
PY - 2026
DA - 2026/09/13
PB - bioRxiv
DO - 10.64898/2026.09.13.751282
UR - https://www.biorxiv.org/content/10.64898/2026.09.13.751282v1
AB - Cognitive workload (CWL) assessment is relevant to a range of applications, such as monitoring driver fatigue and pilot attention, surgeon workload during complex procedures, and astronaut cognitive fatigue during long-duration missions. Eye-tracking datasets are generally small, which hinders the generalizability of the AI models. Time-series foundation models (TSFMs) have shown promise in mitigating this limitation as they are pretrained on large corpora and can be effectively fine-tuned with limited task-specific data. Although cognitive workload measurement has been explored in controlled settings, the ability of TSFMs to generalize to unseen individuals for CWL classification from eye-tracking data has not been studied. In this paper, two TSFMs, MOMENT and Moirai, were evaluated for CWL classification and compared against CNN, FFN, and LSTM baselines on two publicly available eye-tracking datasets. We used subject-level five-fold cross-validation in which data from each test subject were held out during training. We report accuracy, AUC, and F1-score with 95% confidence intervals. On the class-balanced EM-COGLOAD dataset, the pretrained TSFMs generalized markedly better to unseen subjects, with Moirai reaching 0.918 AUC and MOMENT 0.882 AUC, while the task-specific baselines remained near 0.70 AUC. On the class-imbalanced COLET dataset, even though the performance of all the models decreased, MOMENT was the most robust (0.670 AUC). Both TSFMs were also the most stable across held-out subjects, indicating that pretrained representations generalize more consistently across individuals than task-specific models in data-scarce eye-tracking settings.
ER -