Author(s): JOYCE, K., CUTTS, L., ABBOTT, H., STANLEY, B., KANSAKAR, S., SANTOSH, B., BROUGHTON, S., SCOTT, T., DAVIS, S., CLARKE, S., CLARK, T., SIMPKINS, C., GOVES, J., HOWES, E., PATTINSON, K., GARCIA, A., TRENDER, W., ASHDOWN, K. , Institution: UNIVERSITY OF BIRMINGHAM, Country: UNITED KINGDOM, Abstract-ID: 919

INTRODUCTION:
Acute mountain sickness (AMS) affects a large proportion of individuals ascending to high altitude every year. While AMS is not life threatening, if left untreated it can progress to serious illness (e.g., high altitude cerebral oedema). The objective of this study was to conduct a large-scale field study during ascent to Mount Everest Base Camp that measured overnight pulse oximetry among trekkers at several altitudes and further predicted AMS occurring higher up the mountain.
METHODS:
Oximetry-derived desaturation and cardiorespiratory biomarkers, together with clinical metadata, were extracted from participants at Namche, 3443m (n=150; 19.3% AMS-positive) and Dingboche, 4410m (n=130; 20% AMS-positive) and analysed using machine-learning classification. Model performance was compared across five supervised algorithms (logistic regression, decision trees, support vector machine, gradient booster, and random forest) to identify the most predictive features at both locations for AMS occurring at Lobuche (4940m), as determined from clinical assessment performed by the Senior Medical Officer.
RESULTS:
For Namche data, support vector machine (SVM) was the highest performing model, driven by nine key features spanning duration, area, and periodicity-based desaturation biomarkers, along with the age of the participants and prescribed medication status. SVM-Namche exhibited an accuracy of 64 ± 7% (AUC: 0.705±0.082) with 17 of 29 positive cases identified correctly. For Dingboche data, SVM was again the best performing model, driven by a set of features capturing desaturation characteristics (including variability, cumulative burden, and signal pattern measures), overall SpO2 signal distribution properties, median heart rate, and clinically reported headache presence and severity. SVM-Dingboche exhibited an accuracy of 83 ± 7% (AUC: 0.904 ±0.056) with 21 of 26 positive cases identified correctly.
CONCLUSION:
Support vector machines applied to overnight oximetry data (3443m and 4410m) are predictive of AMS occurring at higher altitude days later (4940m) with clinical metadata significantly improving model performance. The predictive power is greater with higher altitudes; however, predictions made at lower altitudes have the advantage of improved response time and may have better clinical outcomes. Further research is required to determine the implementability and readiness of these predictive models.