Five measures from exercise testing helped a model distinguish people who developed acute mountain sickness from those who did not, while also estimating symptom severity. In 330 adult men, the model's AUC, a measure of how well predictions separate groups, ranged from 0.945 to 0.963 across four tasks. The participants had been flown from 500 metres to 3,700 metres within three hours. The study was not designed to show that the fitness measures cause the illness.
How the illness was counted
Researchers used the 1993 and 2018 versions of the Lake Louise Score to assess participants 12 to 24 hours after arrival. Under either version, a total score of at least 3 together with headache counted as acute mountain sickness. The older score classified 126 participants, or 38.2%, with the illness; the newer score classified 113, or 34.2%.
That gave the researchers four prediction tasks: classify acute mountain sickness under each score version and predict the numerical score under each version. They split the 330 participants at the individual level, using 80% for training and holding out 20% for testing. Five-fold cross-validation was kept inside the training group, while the held-out group evaluated performance.
A compact signal from exercise testing
Researchers first compared five candidate models. Random forest had the highest AUC for classifying acute mountain sickness under both Lake Louise definitions. AdaBoost and random forest were the leading models for predicting the 1993 score, while LightGBM led prediction under the 2018 score.
The final ensemble retained random forest, AdaBoost and LightGBM. It used separate copies for each task and combined their predictions with equal weights.
Feature ranking with SHAP, a method that estimates how much each input contributes to a model's output, left five consistently dominant measures: post-exercise heart rate, post-exercise SpO2, AT-RCP time, peak VO2 and ΔVO2 AT-RCP. The five-feature model produced AUCs of 0.947 for classification under the 1993 LLS, 0.948 for classification under the 2018 LLS, 0.963 for predicting the 1993 LLS score and 0.945 for predicting the 2018 LLS score. The reported 95% confidence intervals were 0.862 to 0.997, 0.896 to 0.988, 0.905 to 0.999 and 0.879 to 0.991, respectively.
The signal emerged after exertion
In the old-LLS comparison, the people classified with acute mountain sickness had higher post-exercise HR and VO2/WR slope than those without it. They had lower post-exercise SpO2, AT-RCP time, peak VO2 and ΔVO2 AT-RCP. Each of those listed differences had p < 0.001, while resting SpO2, HR and blood pressure did not differ significantly.
Decision-curve analysis, which tests whether a prediction model offers net benefit at different risk thresholds, reported net benefit above zero across the reported ranges for all four tasks. Incremental net benefit was positive at 10%, 20% and 30% thresholds.
Useful signal, limited reach
The authors also built a web-based support system using the final five-feature model to predict both acute mountain sickness onset and severity under the 1993 and 2018 Lake Louise Score versions.
The evidence has a narrow base. This was a retrospective cohort of young, predominantly Han Chinese men, and the study cannot establish that any of the cardiorespiratory fitness measures causes acute mountain sickness. The proposed mechanism remained theoretical, arterial blood gas and lactate measurements were not collected during exercise testing, and no independent external validation or external calibration was available.
Participants provided written informed consent and the study received institutional human-ethics approval. The work was supported by the China Red Cross Foundation Medical Empowerment Public Welfare Special Fund. The authors declared no commercial or financial conflicts and said raw data supporting the conclusions will be made available without undue reservation.
Paper data and sources
Original title: Baseline indices of cardiorespiratory fitness for predicting acute mountain sickness: an interpretable four-task ensemble model
Authors: Wang S, Miao F, Tao C et al.
Journal/Repository: Frontiers in public health
Status: Peer-reviewed
First online: 2026-10-07
DOI: Not available
Original paper · Full text