WCQMV : Design of A Wavelet Compression Based Quadratic Model for EEG Classification Using Multivariate Analysis

Abstract

Electroencephalogram (EEG) signals represent functioning of the brain, and assist in identification of multiple brain-related disorders including Epilepsy, Alzheimer’s disease, emotional states, Parkinson’s disease, strokes, etc. To design such models, a wide variety of machine learning & deep learning approaches are proposed by researchers. But these approaches use a black-box generic model for EEG classification, due to which their scalability is limited. To enhance this scalability, a novel feature augmented extraction model is proposed in this text. The model uses wavelet compression on input EEG data, and processes the compressed signal using a variance-based selection approach. Due to which, the model is capable of low-delay, and high accuracy classification for different brain-diseases. It evaluates wavelet-based features from input EEG data, and performs ensemble feature selection for improving feature variance. The wavelet features are able to convert input EEG data into different directional components, which assists in improving efficiency of feature representation & model training for different signal types. The proposed model uses a quadratic Neural Network (QNN) classification engine, and is capable of achieving an accuracy of 96.5% for different EEG classes. These classes include 3 types of Epilepsy, presence of Alzheimer’s disease, & evaluation of brain strokes. Due to use of feature variance-based classification, the proposed WCQMV model outperforms existing feature selection & classification models by 4% in terms of accuracy when averaged over multiple datasets. Moreover, the proposed model also improves speed of classification by 4.9% when compared with these models, thus making it useful for high-speed EEG processing applications. This performance improvement is possible due to effective feature reduction, which assists in identification of different EEG signal types. The model was tested on various EEG datasets including, IEEE Port Epileptic dataset, and BNCI dataset for Alzheimer & brain strokes. It was observed that the proposed model was capable of high-performance classification on each dataset, thereby indicating high-scalability across multiple EEG applications.