Performance Analysis of various Neural Network functions for Parkinson’s disease Classification using EEG and EMG

Artificial neural network (ANN) is a significant tool for classification of various types of disease using either Biosignals/images or may be any kind of physical parameters. Establishment of appropriate combination of learning, transfer function and training function is a very tedious task. Here, we compared the performance of different training parameters in feed forward neural network for differentiating of Parkinson’s disease using human brain (Electroencephalogram) and muscle signals (Electromyogram) features as the input vector. 3 different types of training algorithm with six training functions is used. They are Gradient Descent algorithms (traingd, traingdm), Conjugate Gradient algorithms (trainscg, traincgp) and Quasi-Newton algorithms (trainbfg, trainlm). Proposed work compared the mentioned algorithm in terms of mean square error, classification rate (%),R-value and the elapsed time. Study showed that trainlm (Levenberg-Marquardt) best fits for larger data set. It showed the highest accuracy rate of 100% with 0 mismatch classification with a best validation mean square error of 0.0040254 in 3 epochs with a elapsed time of 1.12 seconds. The R-value found was 0.9998 which is in nearly equals to 1. Hence, Levenberg-Marquardt can be used as a training function for the classification of any brain disorder