Comparison of Machine Learning Techniques for Prediction of Diabetes

Diabetes mellitus is a long-term condition characterized by hyperglycemia. It could lead to a variety of problems. According to current trends, the world's diabetes patients will total 642 million by 2040, implying that one in every ten people will be diabetic. Without a doubt, this calls for an immediate action. Machine learning has been applied to many areas of medical health and hence it is also applied to predict diabetes. In this study, Diabetes Mellitus (DM) is predicted by using decision trees, random forests, and neural networks. Physical examination results from a hospital in Luzhou, China, are included in the dataset. It has fourteen characteristics. The models were tested by using fivefold cross validation in this study. Based on the applicability, this study selected several approaches with the best performance to conduct independent test trials. As a training set, 68994 data points were healthy people and 68994 data points were diabetic people. Furthermore, the data considered are almost 5 times at random due to unbalanced data. The result is the average of these five tests. This study has used PCA and mRMR to reduce the dimensionality. When all of the variables are considered, the RF prediction is more accurate: 0.8084.