Physiological markers of low-altitude exercise load for predicting acute mountain sickness: development of an interpretable machine learning model and research web calculator

Abstract

Abstract

Acute mountain sickness (AMS) poses a significant threat to individuals ascending to high-altitude regions, yet early identification of susceptible populations remains challenging. This study aimed to develop an interpretable machine learning model for estimating AMS risk from exercise-induced physiological responses measured at low altitude. A total of 171 healthy participants were included, with 111 participants (51 AMS events) comprising the development cohort and 60 participants (27 AMS events) forming an external validation cohort. Model development was repeated within a nested cross-validation framework, including LASSO feature selection and algorithm selection within the resampling process. The final locked five-feature model included changes in post-exercise pulse oxygen saturation (ΔSpO 2 ) and cardiac output (ΔCO), sex, smoking history, and drinking history, and combined logistic regression, support vector machine, and random forest by equal-weight soft voting. In the development cohort, pooled outer-fold predictions yielded an AUC of 0.866 (95% CI, 0.791–0.929), an AUPRC of 0.842 (95% CI, 0.748–0.932), and a Brier score of 0.148. The calibration intercept was −0.033 (95% CI, −0.504 to 0.527), the calibration slope was 0.931 (95% CI, 0.664–1.351) and the positive net benefit range was 0.078–0.840. In the external validation cohort, the locked model yielded an AUC of 0.834 (95% CI, 0.723–0.929), an AUPRC of 0.867 (95% CI, 0.758–0.946), and a Brier score of 0.152; the calibration intercept was −0.103 (95% CI, −0.834 to 0.531), the calibration slope was 1.013 (95% CI, 0.655–1.666) and a positive net benefit range of 0.200–0.953. Post-hoc SHAP analysis identified ΔSpO 2 and ΔCO as the two most influential predictors. These findings support an association between low-altitude exercise responses and subsequent AMS. However, the prediction model requires further development and validation in substantially larger cohorts before clinical deployment.