Resizing artificial neural networks for automatic detection of epileptiform discharges: A comparison between Principal Component and Linear Discriminant Analysis

Among the all of the different Artificial Intelligence tools, Artificial Neural Networks (ANN) are widely used for automated pattern recognition and classification process. One of its many applications in Biomedical Engineering is its use in the development of accurate and reliable methods for automatic identification of epileptiform discharges in long term electroencephalogram (EEG) recordings. Several methods have been proposed using neural networks as a classifier in which either segments of EEG signal or features extracted from them are presented as ANN input stimuli. Depending on the sampling frequency and signal size, the networks can become quite large, which would affect the performance of a system where this process is inserted. A solution to reduce the dimensionality of the networks is the use of methods such as Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) to reduce the input stimuli and consequently the size of the network. This study aims to analyze and compare the application of PCA and LDA as a tool for dimensionality reduction neural networks that use morphological descriptors to perform the automatic detection of epileptiform discharges. The average efficiency performance of the neural networks with components and LDA selected descriptors as input stimuli was, respectively, 85.58% and 82.28% which is close to the average efficiency (85.46%) achieved with the whole group of descriptors. Throughout the simulations performed in this study we could not detect any apparent relation, direct or otherwise, between the input size reduction and the neural network performance results. However, after simulations we could observe that Principal Component Analysis was better for resizing the neural networks since it had a slightly better efficiency performance at the same time that it reduced the input size in approximately 56%.