Artificial Immune System (AIS)-A Soft Computing Based Approach for Electroencephalography (EEG) Signal Classification

Abstract

Human immune system is characterized as a group of cells, molecules and organs which is capable of performing several tasks, like pattern recognition, learning from stored data in memory, detection of diseases and optimize response against diseases. Development of immunological principles inspired computational techniques are being taken up by the researchers. These techniques are being used to solve engineering problems in the field of artificial intelligence. Extensive research has been undertaken to develop and derive algorithms which are inspired by human immune system. These algorithms use computationally intelligent techniques to model the human system and are known as Artificial Immune Systems (AIS). This research focusses on development of a classification system based on Negative Selection Algorithm (NSA) which uses non-invasive brain electroencephalogram (EEG) recorded with the help of electrodes placed on brain motor cortex. Multi-domain features, time domain and frequency domain, were considered to ascertain the classification accuracy. Mel frequency cepstral coefficients (MFCC) are commonly used as features for audio signal and speech identification. In this research use of MFCC for EEG signal classification demonstrated the highest classification accuracy and selected as the best feature for EEG signals under consideration. Dimensionality reduction is an important aspect of data preprocessing for improving the computational complexity. Stacked auto-encoder, with two pre-trained hidden layers, has been used for EEG data dimensionality reduction. The multivariate motor imagery EEG signals have been classified by set of detectors (artificial lymphocytes) which are trained and optimized using Genetic Algorithm (GA). The underlying rule for training is the negative selection algorithm (NSA), which is developed after taking inspiration from human negative selection principle for maturation of lymphocytes inside thymus. These detector sets are trained and optimized for each class of motor movement for detection of non-self pattern based on a threshold and detector radius. The radius of detector is optimized using GA such that it does not mis-classify the sample of EEG signal. Finally, a comprehensive Negative Selection Classification Algorithm (NSCA) is proposed in this research for classification of brain EEG signals. The AIS based NSCA exhibits improved performance of multivariate classification as compared to the recent techniques used by researchers.