Synchrosqueezing Transform and Non-Negative Matrix Factorization based Feature Extraction from EEG Signals for Motor Imagery Classification

Motor imagery (MI) refers to the mental representation of movement without any motor action. Effective classification of MI tasks is promising for patients with motor disabilities. However, achieving a reliable MI task classification by electroencephalography (EEG) is challenging. The purpose of the study is to improve classification performance by providing discriminative features of intent in MI tasks. In this study, a matrix decomposition technique, namely Non-Negative Matrix Factorization (NMF) is adopted for feature extraction from time-frequency distributions (TFD) obtained by the Waveletbased Synchrosqueezing Transform (WSST). EEG signals were meticulously cleared of noise by Independent Component Analysis (ICA) before performing NMF. Then, the various features such as skewness, kurtosis, discontinuity, standard deviation and sparsity were extracted from NMF vectors in order to train five different classifiers, namely Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Decision Tree (DT), Random Forest (RF) and AdaBoost. WSST provides energy distributions with highly localization capability in TFD, using the reassignment technique. TFDs showed distinctive energy spikes for right-and left-hand MI tasks. The experimental results revealed that the proposed method approaches outstanding accuracy (>95%), kappa (>0.90), and F1 score (0.99) by using KNN, RF, DT and AdaBoost. The study results also show that WSST-based NMF features are promising for MI task classification.