The research team, led by Yasaman Baradaran of the BioSIS Lab at the University of Canberra together with colleagues at Embry-Riddle Aeronautical University, recruited thirty-one healthy young adults aged 19 to 23. Each participant was exposed to four simulated altitude conditions—sea level, 3,000 meters, 4,000 meters, and 5,000 meters—in a computer-controlled normobaric hypoxia chamber, with sessions spaced one week apart and delivered in randomized order. A normobaric chamber reduces the fraction of inspired oxygen at normal barometric pressure, a technique widely used in aviation and military training because it produces physiological effects equivalent to hypobaric exposure while avoiding the operational risks of actual altitude. Peripheral arterial oxygen saturation (SpO2) and heart rate were monitored continuously via fingertip pulse oximetry, and the statistically significant drop in SpO2 across altitude conditions confirmed that the hypoxic challenge was real and dose-dependent.

While seated in the chamber for a 30-minute acclimation period and then throughout testing, participants stood on a static force platform and completed eight standing tasks adapted from the modified Clinical Test of Sensory Integration and Balance protocol. These tasks were deliberately engineered to stress different sensory channels: normal stance, eyes closed, standing on foam, foam with eyes closed, foam with continuous side-to-side head movements to challenge the vestibular system, and the same combination with eyes closed, plus two Sharpened Romberg tandem heel-to-toe conditions with eyes open and closed. Each 30-second trial captured center-of-pressure (CoP) trajectories at 100 Hz, from which the team extracted fifteen sway measures spanning time-domain characteristics such as root-mean-square displacement, path length, sway area, and jerk, and frequency-domain descriptors such as total power, median frequency, and 95% power frequency, computed separately for anteroposterior, mediolateral, and resultant sway directions.

The statistical analysis told a nuanced story. Using Friedman tests with post hoc Wilcoxon signed-rank comparisons and Benjamini-Hochberg false discovery rate correction, the researchers found that no individual sway measure survived multiple-comparison correction—yet the nominal effects clustered strikingly in the foam-based and tandem tasks that place the greatest demands on multisensory integration. The strongest effect appeared in the condition requiring participants to stand on foam while rhythmically moving their heads, where sway total power rose markedly at 4,000 meters compared with both sea level and 5,000 meters. Elevated frequency-domain measures suggested faster, more energetically demanding corrective responses, while increased jerk on foam indicated reduced smoothness and stability. Together, these patterns imply that hypoxic participants needed more frequent and less efficient postural adjustments to stay upright when their sensory inputs were degraded.

Perhaps the most intriguing finding is that the relationship between altitude and sway is not linear. Moderate hypoxia at 3,000 to 4,000 meters increased sway variability and disrupted automatic postural control, but at 5,000 meters sway actually decreased under complex task conditions. The authors interpret this as a shift in motor strategy: moderate hypoxia appears to increase reliance on attentional resources, producing greater variability, whereas more severe hypoxia constrains motor output and suppresses exploratory movement—a stiffening that yields superficially stable but less adaptable control. This constrained strategy may be protective or energy-conserving under severe physiological stress, but it can mask subtle hypoxia-induced changes and reduce the flexibility needed for effective balance in operational settings. The confusion matrix from the machine learning analysis echoed this pattern, with the greatest misclassification occurring between sea level and 5,000 meters, including six sea-level observations predicted as 5,000 meters.

To move beyond the limits of univariate statistics, the team trained four tree-based ensemble classifiers—Random Forest, Gradient Boosting Decision Trees, Extra Trees, and XGBoost—to classify which altitude a participant had been exposed to, using 120 candidate features drawn from the fifteen sway measures across the eight standing conditions. Because the feature count nearly matched the number of observations, the researchers guarded against overfitting with a rigorous design: an 80:20 participant-level split into development and test sets, participant-grouped ten-fold nested cross-validation in which no participant ever appeared in both training and validation partitions, and feature ranking performed separately within every outer training fold so that held-out participants never influenced feature selection. Hyperparameters were tuned with randomized search and inner three-fold cross-validation, evaluating subsets from the single top-ranked feature up to the top thirty.

The optimized Extra Trees classifier achieved the best performance, reaching 70% four-class accuracy using 27 selected features. Extra Trees, or extremely randomized trees, build an ensemble of decision trees with randomized split thresholds, allowing the model to capture nonlinear relationships and complex interactions among sway variables while remaining robust to multicollinearity in high-dimensional data. The feature-importance rankings revealed that the most informative predictors came not from a single task but from a combination of normal stance, foam conditions, and tandem stance—spanning jerk, total power, median and 95% power frequencies in both anteroposterior and mediolateral directions. Notably, some tasks such as normal standing and tandem stance with eyes closed showed no significant effects in the univariate statistical analysis, yet proved sensitive to hypoxia in the multivariate classification. This discrepancy underscores a key methodological insight: individual sway measures may fail to reach statistical significance on their own, but their combined multivariate patterns still carry meaningful information about altitude-induced functional impairment.

The implications for aviation safety are substantial. Spatial disorientation remains a persistent contributor to pilot fatalities, and hypoxia is a known aggravating factor, particularly when visual references are degraded. Because postural control reflects the coordinated function of visual, vestibular, and somatosensory systems, objective sway measurement may serve as a sensitive early marker of hypoxic impairment—potentially flagging degradation at altitudes thousands of meters below where aviation regulators currently require supplemental oxygen. The authors suggest that integrating postural monitoring into operational safety systems could enhance early warning capabilities during unpressurized flight, cabin depressurization, or periods of high workload, informing decisions about supplemental oxygen use before perceptual symptoms such as color vision loss ever appear.

The study is not without limitations. The sample was modest and restricted to healthy young adults, limiting generalizability across age, sex, and occupational populations. The normobaric exposure paradigm, while physiologically equivalent in many respects, may not fully replicate hypobaric hypoxia encountered at true altitude, and the cross-sectional design could not capture adaptation or acclimatization effects across repeated exposures. The authors also note that model-agnostic interpretability tools such as SHAP values were not applied, and that conventional scalar sway features may miss the nonlinear temporal dynamics of postural control that fluctuation and recurrence quantification analyses could reveal. Future work with larger and more diverse cohorts, longitudinal designs, external validation, and integration with neurophysiological measures such as functional near-infrared spectroscopy or brain imaging will be needed before these findings can inform aviation procedures or oxygen requirements.

Even so, the study establishes a compelling proof of concept: hypoxia acts on the balance system in a multidimensional, stage-dependent way that only becomes fully visible when statistics and machine learning are used in tandem. By treating postural sway as a high-dimensional signature rather than a handful of scalar metrics, the approach detects subtle physiological changes that traditional group-level comparisons overlook. As machine learning continues to penetrate aerospace and biomedical monitoring, the humble force platform may become an unexpected sentinel—catching the silent onset of hypoxia in the wobble of a standing body long before the mind registers anything is wrong.

Subject of Research: Effects of hypoxic altitude exposure on postural stability assessed with statistical analysis and machine learning classification of postural sway

Article Title: Integrating machine learning and statistical techniques to evaluate postural stability during hypoxia

Article References: Baradaran, Y., Shrestha, A., French, J., & Ghahramani, M. (2026). Integrating machine learning and statistical techniques to evaluate postural stability during hypoxia. Frontiers in Bioengineering and Biotechnology, 14, Article 1886164. https://doi.org/10.3389/fbioe.2026.1886164

Image Credits: AI Generated

DOI: 10.3389/fbioe.2026.1886164

Keywords: hypoxia, postural stability, machine learning, Extra Trees, center of pressure, spatial disorientation, aviation safety, normobaric chamber, sensory integration, feature selection, XGBoost, aerospace medicine