EEG motor imagery classification using machine learning techniques

Abstract

A brain-machine interface (BMI), is a device or experimental setup that receives a brain signal, classifies it, and then uses it as a computer command. Even if large amounts of work exist in the field, there is not a consensus on which kind of learning methodology (deep learning, convolutional networks, AI, etc.) and/or type of algorithms in each methodology, are best to run BMIs. The aim of this work was to build a low-cost, portable, easy-to-use and a reliable BMI based on Motor Imagery Electro-encephalography. To this end, different algorithms were compared to find the one that best satisfied such conditions. In this study, motor imagery EEG signals, from both PhysioNet public data and from our own laboratory obtained using an Emotiv headset, were classified with four machine learning algorithms. These algorithms were: Common spatial patterns combined with linear discriminant analysis, Deep neural network, convolutional neural network and finally Riemannian minimum distance to mean. The mean accuracy for each method was 78%, 66%, 60% and 80% respectively. The best results were obtained for the baseline vs Motor Imagery comparison. With global-training public data, an accuracy between 86.4% and 99.9% was achieved. With global-training lab data, the accuracy was above 99% for Common Spatial Patterns and Riemannian cases. For lab data, the classification/prediction computing time per event were 8.3 ms, 18.1 ms, 62 ms and 9.9 ms, respectively. In the discussion a comparison between the results presented here and state-of-the-art of methodologies and algorithms for BMIs can be found. We concluded that Common spatial patterns and Riemannian minimum distance to mean, algorithms resulted in fast (computing time) and effective (success rate) tools for their implementation as deep learning algorithms in BMIs.