Comparison of Artificial Neural Network, Logistic Regression and Discriminant Analysis Efficiency in Determining Risk Factors of Type 2 Diabetes

Introduction: Diabetes is the most common endocrine disease caused by sugar, fat and protein metabolism disorder and is characterized by blood sugar increase. Pre-diabetic individuals are the most vulnerable people at risk of diabetes, therefore; in the present study pre-diabetic individuals are considered as control group versus diabetic patients. Depending on the nature of dependent and predictor variables, there are various statistical models that suit various situations to recognize and classify these characteristics. To achieve the above propose, this study analyzes efficiency and prediction power of three statistical models. Materials and Methods: Data are collected from 17 rural health centers in Kermanshah city. An experimental group of 100 diabetic and a control group of 100 pre-diabetic patients were entered into the study. The under study variables included demographic data, body mass index, fasting blood sugar, glucose tolerance, blood pressure, blood lipid and individuals’ daily activity. The data were recorded in 2 separate checklists from the subjects’ latest health record data obtained from Kermanshah rural health centers. Artificial Neural Network (ANN), logistic regression and discriminant analysis models were applied for data processing to identify risk factors. ROC curve was used to compare prediction powers of the models. To specify a model with the highest prediction, Radial Basis Function (RBF) and wrapper method were applied in ANN model. The method which took all the situations into consideration was applied to enter independent variables to the model; subsequently, a model with the highest prediction power was selected as ANN superior model. Results: According to area under the ROC curve, prediction power of the three models; RBF, logistic regression and discriminant analysis models were estimated as much as 0.864, 0.884 and 0.80, respectively. Gender variables (P=0.027) and fasting blood sugar (P<0.001) in Logistic regression model and age variables (P=0.014), fasting blood pressure (P<0.001) and glucose tolerance (P<0.001) in discriminant analysis model indicated significant correlation. According to wrapper method; the model consists of fasting blood sugar, glucose tolerance, BMI and activity with 82.1% prediction power turned out to be selected as the best RBF pattern (out of 242 possible models). RBF with 95.2% indicate the highest sensitivity among the three models. Conclusion: At the present study RBF had a higher level of accuracy and sensitivity although logistic regression performed more powerful to distinguish between diabetic and pre-diabetic patients. In communities with high affinity between experimental and control groups demand stronger methods to discover the differences between the groups. Therefore, application of these methods in medical studies is recommended.