Machine Learning for Detection and Classification of Motor Imagery in Electroencephalographic Signals

Patients with motor disabilities usually have high intellectual capacity, but with great physical dependence. Despite this, brain signals are generated in regions of the cortex when movements are imagined. These signals can be translated using motor imagery. Motor Imagery is highly complex due to the variability and non-stationary nature of the signals. The use of machine learning is fundamental for detecting and classifying these movements. In this chapter, we experimented simple and quick classifiers for recognizing left and right hand movements. The study was carried out using the 2b database of BCI Competition 2008. For this, signal processing techniques were tested, including bandpass filter, wavelet decomposition and statistical thresholds in wavelet components. Then, numerical attributes were extracted in the time and frequency domains. All experiments were performed 30 times with 10 folds cross validation. Furthermore, the SMOTE method was applied to the collected signals. With this method, we simulated scenarios where more signals were collected from patients. The results showed a better classification performance with the raw data collected. This indicates that important information was lost in processing and that other settings should be tested. In addition, there was a significant improvement in results with the expanded database. Therefore, this work suggests that the database is insufficient for generalization.