Detecting movement thoughts from electrodes in the brain

Abstract

Brain-computer interface (BCI) systems enable direct communication from the brain to an output device, bypassing the traditional pathway of peripheral nerves and muscles. These systems can record brain signals from either the scalp, the surface of the cortex, or from within the brain and allow users to control numerous applications such as neural prosthetics and computer programs.

The data used in this project are brain signal recordings obtained by means of an invasive dataacquisition method known as stereo-electroencephalography (sEEG). The data was collected from two patients who were opening and closing two types of pliers: normal and reverse pliers. The hand movements used to manipulate the pliers create four phases of plier tip movement, namely the ‘opening’, ‘fully open’, ‘closing’ and ‘fully closed’ phases. This project makes use of the sEEG signals to determine the intended hand movements by classifying the phases and to determine the reliability of the feature extraction techniques. One type of feature extraction technique which is used is based on spectral analysis. The other is based on a popular technique called common spatial patterns (CSP), which provides a set of spatial filters that optimally discriminate between two classes of data in the least-squares sense.

The results show that good performance is achieved when classifying pairs of phases using features extracted through spectral analysis, with classification accuracies lying between 63.4% to 95.6% across the pairs. Therefore, these features are a good representation of the classes of hand movements. Classifying the four movements using these features gave an average classification accuracy of 48.1%. The classification performance between pairs of phases using features derived from the CSP algorithm is poor, with accuracies ranging from 40.1% to 68.8% across the pairs. Classifying the four movements resulted in an average classification accuracy of 31.2%. The CSP algorithm appears to suffer from overfitting and poor generalisation on sEEG data.