Abstract
BACKGROUND: Sarcopenic obesity (SO) has a higher risk of adverse health events compared to having obesity or sarcopenia alone due to the dual burden of both muscle loss and fat gain. The prevalence of SO is progressively increasing as the population ages and the obesity epidemic progresses. Currently, there are no tools for predicting the risk of SO, and this study aimed to construct a well-performing prediction tool based on machine learning. METHODS: The National Health and Nutritional Examination Surveys (NHANES) 1999-2004 dataset was used for the analysis, and the included data were randomly divided into training and validation sets in the ratio of 70:30. Missing data is processed using multiple interpolation technique. A 5-fold cross-validated recursive feature elimination algorithm is used to rank the importance of features, and the top three important features from each algorithm are used as the features of the model for constructing the machine learning model. Six machine learning methods, CART, GBM, KNN, LR, NNet, XGBoost, were used to develop models and evaluated for discrimination, calibration, clinical utility, and robustness. The combined best-performing model was further developed into an online web calculator for clinical applications. RESULTS: The study had 5607 participants, and 1139 (20.3%) of them had SO, with a prevalence of 21.2% among females and 19.4% among males. Combining all the performance evaluations, the GBM-based model has the best performance, which uses age, race, and BMI as the features of the model, and its AUC values in the training and validation sets are 0.820 and 0.832, and it has good calibration, clinical utility, and robustness. CONCLUSION: In this study, the GBM-based model performed well, and an online web calculator constructed on the basis of the model was used to identify the risk of SO in the US community for those over the age of 60.