Time and Frequency Domain Features Extraction Comparison for Motor Imagery Detection

The brain computer interface area has increased the number of applications in the last years, searching to improve the quality of life in injured people. In spite of the progress in the field, different strategies are analyzed in order to contribute in specific problems related to the main applications. Present proposal shows a comparison between the use of time or frequency domain for feature extraction in upper limbs motor imagery. Four machine learning techniques as K-Nearest Neighbor, Support Vector Machine, Neural Networks and Random Forest were trained to detect motor imagery from EEG signals. Comparison for feature extraction and the employed detection models were analyzed to find the best election in an application for close-open fist in hands for two scenarios, according to two or three classes classification. The results achieved more than 90% in accuracy for both domain approaches in the two classes case. For the three classes detection, the results dropped out to 87% in accuracy. In general, the frequency domain is preferable for feature extraction and the KNN classifier was the best strategy for the present study.