Combined Spatial Filtering and Time-Frequency Analysis in Motor Imagery BCI

Abstract

Brain computer interfaces (BCI) provide ways for the brain to communicate directly with a computer, and are today primarily used as an aid for people with severe movement impairing disorders. In this study a BCI feature extraction method using a combination of spatial filtering and time-frequency (TF) analysis is developed. The features are then classified using a convolutional neural network. A four class motor-imagery data set, data set 2a from BCI competition IV, is used for training and evaluation.

The spatial filtering is performed using the common spatial pattern (CSP) algorithm, and three different time-frequency analysis methods are implemented: the spectrogram based on the short-time Fourier transform, the Wigner-Ville distribution and the wavelet transform. The methods are evaluated using the Cohen's kappa performance measure, and the results are compared.

Combining spatial filtering and TF analysis significantly improves the classification results, as compared with any of the stand-alone methods. The best result is achieved by the combined CSP and spectrogram method, yielding a kappa of 0.621. The combined CSP and wavelet method performed similarly, achieving a kappa of 0.620.

Popular Abstract

Providing means of communication for paralyzed individuals is an important and fascinating subject. The most famous person using one of these methods was probably Stephen Hawking. He used a program that tracked his cheek movements, allowing him to move a cursor on a screen and in turn form words. As advanced as that might have been at the time, there are cases where this is not enough. If the patient has no control of their muscles at all this method also becomes unavailable. To provide a means of communication for people with no control over their body, one has to stop relying on the muscles themselves as the medium of communication, and instead use the signals of the brain directly.

Brain computer interfaces (BCI) are, as the name states, ways for the brain and a computer to communicate, by using the brain waves as signals. Often this works by the patient imagining a movement, for example raising their hand. Sensors reading the brain waves associated with the movement in turn translate this to a signal for the computer. This is called motor imagery (MI), and is the type of data considered in this project.

While BCI might sound straightforward, there are several problems associated with reading and then classifying the brainwaves. There are methods which are very good at reading brain activity, but are too unpractical to be used for everyday communication. A person cannot be expected to lie in a Magnetic Resonance Imaging (MRI) machine every time they want to say a word. On the other end of the spectrum are methods which are very easy to use, but in turn produce less clear data. Among the most common of these methods is electroencephalography (EEG), consisting of placing electrodes on the patient's scalp and measuring the voltage on the different cranial positions. This is very easy to use in practice, the limiting factor being the possession of a fancy cap to measure the brainwaves. Sadly the measurements are not only of the imagined movement, but the EEG also pick up many other disturbances, such as eye movement and blinking. While unrelated muscle movement is not really an issue for those with complete lack of muscle control, it most certainly is for everyone else using it. Taking into consideration that the training data for the classifiers are mostly collected from healthy individuals, this creates the need for the classifier to on its own filter out the noise and only keep the relevant parts of the signals.

In this project a new method for interpreting brain activity was developed. The method combines spatial filtering and time-frequency analysis. The spatial filtering allows the BCI to focus on certain brain regions that are of high interest. This is done using the machine learning algorithm called the Common Spatial Pattern, which is very prominent in the field of BCI. The time-frequency analysis is then used to transform the signal into a form which simplifies the interpretation of the oscillatory characteristics of the brain activity.

The combination of the different methods provide substantially better results than any of the methods do alone, and show that spatial filtering may be a key component in BCI systems utilizing time-frequency analysis.

• author: Welander, Andreas and Robertsson, Max
• supervisor: Rachele Anderson
• organization: Mathematical Statistics
• course: FMSM01 20202
• year: 2021
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics
• keywords: Brain Computer Interface, Machine Learning, Convolutional Neural Network, Time-Frequency Analysis, Wavelet Transform, Wigner-Ville Distribution, Spectrogram.
• publication/series: Master's Theses in Mathematical Sciences
• report number: LUTFMS-3411-2021
• ISSN: 1404-6342
• other publication id: 2021:E12
• language: English
• id: 9042769
• date added to LUP: 2021-05-12 09:42:16
• date last changed: 2021-06-03 15:20:33