Multi-Level Counter Propagation Network for diabetes classification

The data mining techniques can be successfully used for the classification of patients suffering from diabetes. The classification can be done to find out categories such as not detected, initial stage, middle stage and advanced stage of diabetes. This study has undertaken only two class based classification of positive (diabetes detected) and negative (diabetes not detected) class. To perform the classification using data mining techniques, the input data are used from Pima Indian Diabetes (PID) sample data sets which is available as an open source. Three classification techniques of Radial Basis Function (RBF), Multi Layer Perceptron (MLP) and Multi Level Counter Propagation Network (MLCPN) are used to classify diabetes cased. The MLCPN functions in two phases, the first phase is Kohonen Self Organizing Map (KSOM) and second phase is Grossberg learning. Both the methods together make hybrid approach. Among the models analysed, the MLCPN produced better accuracy and efficiency. It was faster and an accuracy rate was approximately 97%. The simulation tests were performed using Weka software tool. A total of 519 datasets were used for training the models and as many remaining dataset were used for testing. Simulation results were on the lines of expectation with good accuracy.