Background: Diabetes mellitus as a chronic disease is the most common disease caused by metabolic disorders and it is one of the most important health issues all around the world. Nowadays, data mining methods are applied in different fields of sciences due to data mining methods capability. Therefore, in this study, we compared the efficiency of data mining methods in predicting type 2 diabetes.
Methods: In this cross-sectional study, the data of 7,000 participants in the Diabetes Screening Project in Samen, Mashhad City, Iran, were considered in 2016. There were 540 untreated diabetic patients. The Samen Project was included in the routine examinations of diabetes patients like blood glucose, eyes health, nephropathy, and legs health. So, in order to maintain balance, 600 healthy individuals were selected in a proportional volume sampling in this study. Therefore, the total sample size was 1140 people. In this study, people with diabetes aged over 30 years old were enrolled and participants with the previous history of type 2 diabetes, with normal blood glucose due to drug use or other issues at the time of the study, were excluded.
Results: All three models (Logistic regression, simple Bayesian and support vector machine models) had the same test accuracy (86%), however, in terms of area under the receiver operating characteristic (ROC) curve (AUC), logistic regression and simple Bayesian models had better performance (AUC=90% against AUC=88%). In the simple Bayesian model and logistic regression, body mass index (BMI) and age variables were the most important variables, while BMI and blood pressure variables were the most important factors in the support vector machine model.
Conclusion: According to the results, all three models had the same accuracy. In terms of area under the curve (AUC), logistic and simple Bayes models had better performance than the support vector machine model. Totally all three models had almost the same performance. Based on all three models, BMI was the most important variable.