Abstract: Objective To put forward the idea of constructing an early warning model for college students' physique， help college students to make reasonable predictions on each physical test item level by comparing the performance effects of different models in forecasting， so as to achieve the purpose of early warning. Methods On the basis of collecting the 2015-2018 physical measurement data of Nanjing University of Aeronautics and Astronautics， using data mining technology， after data understanding and data preparation， three machine learning algorithms such as Random Forest Algorithm， Gradient Boosting Decision Tree Algorithm and Neural Network Algorithm were used to build the model and evaluation. Results From the perspective of the prediction effect on the final physical grade of boys， the accuracy rate was between 90% and 97%，and the accuracy rate of the Gradient Boosting Decision Tree method was always better than the other two. When the training set ratio was 80%，the highest prediction rate reached 96.19%. From the perspective of predicting the final fitness level of girls，the accuracy rates of Random Forest and Gradient Boosting Decision Tree methods were between 90% and 96%，while the accuracy rates of Neural Networks methods were between 80% and 87% . When the training set ratio was 80%， the Gradient Boosting Decision Tree method achieved the highest accuracy rate of 95.06%. Conclusion It is feasible to construct an early warning model for college students' physique，and the prediction accuracy of the final physique grade can reach more than 93%， of which the Gradient Boosting Decision Tree method has the best performance.

Key words: Physique health, Data mining, Feature engineering, Machine learning