Research On Diabetes Prediction Models Based On Machine Learning Algorithm

Currently, the number of patients of chronic diseases in China is the most one in the world, and diabetes and its related complications are important component of chronic diseases. Residents have strong health needs. Therefore, it is important to build diabetes prediction model to estimate risk of diabetes for people and find high-risk groups, then gives then early warning of diabetes.After learning formal research results and analyzing the diabetes risk factors,this research do stepwise regression analysis on the data set, in order to remain the most significant characteristics components as the input variables of predicition models based on BP Artificial Neural Network, SVM Theory and Ensemble Learning.Machine learning algorithms are good at dealing with complex issues because their high accuracy and generalization ability. 2728 sample set of data are devided in accordance with 7: 2: 1 ratio into one training set,one test set and an independent sample set. BP artificial neural network, support vector machine learning model,ensemble learning model are established. Input variables, model parameters, and the choice of kernel are more or less influence on the prediction results. We observed the influence of the change of network structure, learning rate, the punishment factor, kernel function and its parameters on prediction results, and to adjust to the best case, then select the best model of each algorithm.Finally, we use independent samples for testing, the results of three prediction models have a high degree of correlation with raw data, which proved to be statistically significant. In three models, artificial neural network model achieved a higher AUC and less time. Finally, the article chose artificial neural network model with 7-1-1 network structure is the optimum prediction model for diabetes prediction.