Selection of Artificial Neural Network Training Algorithms in the Detection and Classification of Wavelet de-noised Musical Tone Stimulated EEG Signals
Authors: Roy Francis Navea and Elmer Dadios
Abstract
The human brain can be stimulated by internal and external factors with which the effect of these can be traced from brainwaves or EEG signals. The natural complexity of EEG signals calls for methods by which information can be extracted and used for a particular purpose. In this study, musical tones were used to stimulate the brain and an attempt was made to detect and classify these stimulations from the EEG signals. An Artificial Neural Network (ANN)-based classifier was employed to do this task. Wavelet based de-noising was used to smoothen the musical tone stimulated EEG signals and among the 110 known mother wavelets, the reverse biorthogonal ‘rbio3.1’ and ‘rbio3.3’ using the ‘rigrsure’ thresholding method satisfied the selection criteria for better denoising effects.
Detection and classification were performed using ANNs implementing four different training algorithms. Results show that trainbr or trainlm is good for detection while the trainlm was found to be better than the other training algorithms used when it comes to classification. The metrics for selecting the training algorithm were based on the F-score and the rejection rate having the condition that F-score should be high while the rejection rate should be low.