Model Development and Validation for Classifying Hypoxia in Military Aircrew Using ECG and Skin Temperature

Background: Hypoxia occurs when blood or tissues are deprived of adequate oxygen, posing a significant risk to military aircrew operating at high altitudes due to reduced atmospheric pressure. The danger lies in its subtle symptoms—such as impaired judgment—often unnoticed until serious consequences arise. While hypoxia can be detected using direct (e.g., pulse oximetry), indirect (e.g., PPG), or tissue-level (e.g., NIRS) methods, these are often impractical in flight settings due to motion artifacts, low perfusion, or invasiveness. This study aims to develop and internally validate machine learning models to classify hypoxic conditions in military aircrew members using electrocardiogram (ECG) and skin temperature signals, offering a non-invasive and real-time monitoring approach. Methods: Data were collected from healthy military aircrew members undergoing standardized hypoxia training in a hypobaric chamber simulating high-altitude conditions. ECG, skin temperature, and respiration signals were recorded using wearable sensors. Hypoxia events were labeled based on oxygen mask removal at altitude. A multi-window feature extraction approach was applied using time windows of 20–120 seconds, enabling trend detection across scales. In total, 87 ECG-derived features were combined with temperature and respiration features. After preprocessing and quality control, a feature selection strategy based on repeated classifier rankings was employed. Five classifiers were trained and evaluated (Support Vector Classification, Decision Tree, Random Forest, Logistic Regression, K-Nearest Neighbors) using cross validation and accuracy and log loss metrics. Results: Data from 40 participants were included after preprocessing. Across classifiers, classification accuracy ranged from 0.85 to 0.90, with the SVC achieving the highest average accuracy (0.90 ± 0.07) and lowest log loss (0.25 ± 0.15). The most informative Online First features came from longer time windows, particularly the 80th percentile of respiratory rate intervals (HRV_Prc80NN), mean respiration rate, and mean skin temperature. Classifier performance was robust across models, with small differences, suggesting that model architecture is less critical than feature representation. Conclusions: We demonstrate that ECG and skin temperature signals can reliably detect hypoxic conditions in military aircrew using machine learning. The approach shows strong internal performance and highlights specific physiological features as key indicators. These findings support the feasibility of real-time, non-invasive hypoxia monitoring in flight environments and lay the groundwork for future applications in other high-risk domains such as commercial aviation, spaceflight, and clinical monitoring. Further research will involve evaluating model performance in operational flight conditions and exploring generalization across individuals and sensor systems.