Construction and Risk Factor Analysis of the High-altitude De-acclimatization Syndrome Prediction Model Based on Stacking Ensemble Learning

ZUO Jinxin; WANG Rui; LUO Yongjun

doi:10.13730/j.issn.2097-2148.2025.09.010

Abstract：Objective To identify risk factors for high-altitude de-acclimatization reaction and to develop the risk-prediction model to support targeted prevention.Methods In August 2022,1 544 individuals who descended from the Ti-betan plateau(≥3 000 m)to plain were recruited by cluster sampling,the data were collected through the questionnaire survey,and the subjects were divided into the high-altitude de-acclimatization reaction group(n=192)and the non-high-altitude de-acclimatization reaction group(n=1 352).Decision tree(DT),random forest(RF)and extreme gradient boosting(XGBoost)served as base model,Logistic regression was used to construct the Stacking ensemble learning mod-el,the receiver operating characteristic(ROC)curve area under the curve(AUC)of 4 single prediction models and Stac-king ensemble learning models were compared,and the predictive efficacy of the model was evaluated.In this study,70％of the questionnaire data of high-altitude de-acclimatization reaction were randomly divided into the training set and the re-maining 30％into the test set.According to the AUC value of the model test set,the importance of the characteristics of the variables screened by each model was weighted average,and a comprehensive importance rank-ing of the variables was obtained to find the risk factors for the occurrence of high-altitude de-acclimatization reac-tion.Results Among the 1 544 subjects,192 had high-altitude de-acclimatization reaction.The AUC values of Logistic regression,DT,RF,XGBoost and Stacking ensemble learning models in the test set were 0.757,0.601,0.788,0.793 and 0.853,respectively.DT,RF and XGBoost models comprehensively screened out the risk factors of high-altitude de-acclimatization reaction,including the number of days that had been to plateau,acute mountain sickness(AMS)of this time,family history of chronic diseases,altitude of plain residence,smoking,means of transportation back to the plain and body mass index(BMI).Conclusion Taking the way of transportation back to the plain,the number of days in the history of visiting the plateau,AMS situation of this time,the altitude of residence,smoking,family history of chronic diseases and BMI as predictors,the Stacking ensemble learning model constructed had the best comprehensive perform-ance in the test set,which was convenient for application.