Prediction of diabetes by Neural Network

Meta-heuristic and combined algorithms have a great capability in modelling medical decision making. This study used neural networks in order to predict Type 2 Diabetes (T2D) among high risk individuals.
This study was an applied research. Data from 545 individuals (diabetic and non-diabetic), in Diabetes Clinic of Hamedan University of Medical Sciences, were used to develop predictive diabetes models. Memetic algorithms which are a combination of genetic algorithm (GA), local search algorithm, and neural networks were applied to update weights and improve predictive accuracy of neural network models. In the first step, optimum parameters for neural networks such as momentum rate, transfer functions, and error functions were examined through trial and error and other studies.
The preliminary analysis showed that the accuracy of neural networks was 88 percent. The use of memetic algorithm improved its accuracy to 93.2 percent. Among models, regression model had the least accuracy. For the memetic algorithm model the amount of sensitivity, specificity, positive predictive value, negative predictive value, and area under the curve were 96.2, 95.3, 93.8, 92.4, and 0.958, respectively. These parameters for GA were 98.0, 84.8, 88.6, 98.2, and 0.952 and for the logistic regression model were 95.6, 84.5, 94.7, 87.0, and 0.916, respectively.
Models developed by neural networks have a higher predictive accuracy than the regression model. The results of this study can contribute to risk management and planning of health services by providing healthcare decision makers with more accurate predictive models based on clinical and life style characteristics of individuals.