Objective To explore the value of CatBoost model in predicting severe hand-foot-mouth disease (HFMD) by the machine learning algorithm. Methods A total of 2 983 children with HFMD diagnosed and treated in a hospital in Zhengzhou from January 2014 to June 2017 were collected, data were analyzed with R 3.4.3 software, CatBoost model and other common models were constructed, prediction performance of CatBoost model was evaluated. Results The predictive accuracy of the finally constructed CatBoost model was 87.6%, artificial neural network model ranked second (83.8%), other models (decision tree, support vector machine, logistic regression, Bayesian network) had predictive accuracy less than 80%. The area under receiver operating characteristic (ROC) curve, sensitivity, and specificity of CatBoost algorithm model were all high (0.866, 80.80% and 92.33% respectively), the top three predictive variables were vomiting, limb jitter, and pathogenic results. Conclusion CatBoost model can be used to predict severe HFMD, which has higher accuracy and diagnostic value than other traditional algorithms.