Machine learning models for mental workload classification trained under normal conditions show severely degraded performance under hypoxia and/or sleep restriction, but models trained across all physiological conditions using multimodal EEG and eye tracking data maintain robust classification performance.
Key Findings
Results
The best machine learning classifiers for mental workload classification under habitual sleep and normoxia conditions were K-Nearest Neighbors, Support Vector Machine, and Random Forest.
KNN achieved an F1 score of 80.3 ± 8.9%, SVM achieved 77.8 ± 10.3%, and RF achieved 75.7 ± 9.1%
Classification was performed on 1-minute MW level epochs across three levels (low, medium, high)
Features were selected from EEG, ECG, respiratory, and eye tracking sensors using backward Recursive Feature Elimination (RFE)
Six machine learning classifiers in total were compared in this study
Results
Exposure to sleep restriction and/or hypoxia substantially degraded mental workload model performance when models were trained only on normal physiological conditions.
F1 scores dropped to below 35% when models trained on habitual sleep/normoxia (HSNO) were applied under sleep restriction and/or hypoxia conditions
Sleep restriction was defined as less than 3 hours Total Sleep Time, while habitual sleep was defined as more than 6 hours TST
Hypoxia was induced at FIO2 = 13.6% for 2 hours, simulating approximately 3500 m altitude
This represents a substantial decline from baseline F1 scores of approximately 75–80% under normal conditions
Results
Models trained on data from all physiological conditions combined (All-Conditions) maintained robust mental workload classification performance across conditions.
KNN trained on All-Conditions achieved F1 scores of 77.4 ± 7.8% across conditions
RF trained on All-Conditions achieved F1 scores of 70.7 ± 10.2% across conditions
Models including EEG and eye tracking features performed particularly well
These All-Conditions models preserved performance comparable to the best single-condition models
Methods
The study used a 4-condition crossover design with 17 healthy participants exposed to combinations of sleep restriction and hypoxia during a mental workload task.
Seventeen healthy participants were randomly exposed to all four conditions: habitual sleep/normoxia (HSNO), habitual sleep/hypoxia (HSHY), sleep restriction/normoxia (SRNO), and sleep restriction/hypoxia (SRHY)
Mental workload was manipulated using the Multi-Attribute Test Battery (MATB)-II with an additional auditory Oddball-like task
Three 12-minute periods of increased MW (low, medium, and high) were used per condition
The trial was registered under NCT05563688
Results
Multimodal physiological data including EEG and eye tracking were identified as critical sensor modalities for robust cross-condition mental workload classification.
Features were derived from EEG, ECG, respiratory, and eye tracking sensors
Models specifically including EEG and eye tracking showed the best cross-condition generalization
Feature selection was performed using backward Recursive Feature Elimination (RFE)
The study context was aircraft pilot operations, where pilots may simultaneously face high mental workload, hypoxia, and sleep restriction
Background
The study was motivated by the operational reality that aircraft pilots can simultaneously experience high mental workload combined with moderate hypoxia and sleep restriction.
The study simulated conditions relevant to aviation environments where physiological stressors co-occur with demanding cognitive tasks
Hypoxia level of approximately 3500 m was chosen to reflect moderate altitude exposure relevant to aviation
The authors aimed to assess cross-validation of MW predictive models under these combined stressors
A secondary aim was to develop a robust predictive model valid across different physiological conditions
What This Means
This research suggests that artificial intelligence systems designed to monitor pilots' mental workload in real time can fail dramatically when pilots are also experiencing reduced oxygen (hypoxia) or lack of sleep — two conditions that are common in aviation. The study tested 17 participants on mental workload tasks under four combinations of normal sleep vs. sleep deprivation and normal oxygen vs. reduced oxygen levels, using brain activity (EEG), heart rate, breathing, and eye tracking sensors to train machine learning models. When models were trained only on data from well-rested, well-oxygenated participants, their ability to correctly classify mental workload dropped from around 75–80% accuracy to below 35% when participants were sleep-deprived or hypoxic.
This research suggests that the solution lies in training these AI models on data collected under multiple physiological states, not just ideal conditions. Models trained on data from all four conditions — including sleep restriction and hypoxia — maintained strong performance (around 70–77% accuracy) regardless of which condition a participant was in. Brain wave (EEG) and eye tracking data were found to be particularly important for this robust performance across conditions.
For practical applications in aviation safety, this research suggests that mental workload monitoring systems should be developed and validated under the full range of physiological conditions pilots actually encounter, rather than only under controlled laboratory conditions. Relying on models trained only in ideal conditions could give false reassurance about a pilot's cognitive state when they are most vulnerable to performance degradation.
Pontiggia A, Quiquempoix M, Fabries P, Beauchamps V, Jacques C, Guillard M, et al.. (2026). Robust multimodal mental workload classification: A cross-physiological condition machine learning approach.. Computer methods and programs in biomedicine. https://doi.org/10.1016/j.cmpb.2026.109251