Predicting Open/Closed Fist Motions from EEG Data Using Machine Learning

Brain-computer interfaces, or BCIs, are systems that use neural signals from the brain in order to produce a certain action. There is so much potential with BCIs, but because of the infancy of the technology current research is mostly limited to medical solutions. One of these solutions is to control prosthetic devices. I decided to focus my research on this prosthetic hand manipulation through the use of electroencephalography. This system is non-invasive and allows for the collection of the electrical activity of the brain via electrodes placed on the surface of a participant’s scalp. The purpose of my research was to create a predictive algorithm that takes an input of neural signals and outputs whether the user would like to open or close their fist. I used an EEG headset with 14 scalp electrodes called the EMOTIV EPOC X. Data collection proceeded with one participant and 50 blocks of 4 randomized trials. The trials were as follows: physically closing one’s fists, physically opening one’s fists, imagining closing one’s fists, and imagining opening one’s fists. The data was collected in the EMOTIV Pro software and exported to MATLAB for processing. I used a band-pass filter to extract the beta frequency band for the signals correlating to physical movement, and the mu frequency band for the imagined movement signals. For my predictive algorithm I used machine learning. Specifically, I used an artificial neural network where I split my data 80% towards training and 20% towards testing. The results were that my algorithm had a 41.8% accuracy in predicting the physical opening and closing of one’s fists and a 58.6% accuracy in predicting the imaginative opening and closing of one’s fists. Therefore, there was a low accuracy in my algorithm’s ability to predict fist motions based on EEG data.

Hi, My name is Ayman Slamani and the research project that I worked on this semester was “Predicting Open/Closed Fist Motions from EEG Data Using Machine Learning”. This project is under the Bioengineering department, and my mentor was Dr. Siddhartha Sikdar.

I want you to imagine a world where you can control every device you own with just a thought. This world is not a fantasy and can come to fruition if we can better learn how to read, process, and decode our neural signals. Brain-computer interfaces, or BCIs, are systems that literally connect a user’s brain with a computer and use neural signals from the brain in order to produce a certain reaction to the data that is analyzed by the computer. There is so much potential with BCIs, but because of the infancy of the technology current research is mostly limited to medical solutions. One of these solutions is to control prosthetic devices. Because it is important to have clean and precise data to be able to manipulate external devices with, most research to do with BCIs use invasive methods to collect their data. This means that participants would have to go through neurosurgery, and have electrodes implanted into the surface of their brain to be able to send accurate data to the computer. Because of the risk that surgeries like this can have, I decided to focus my research on the use of electroencephalography or EEG. This method is non-invasive and allows for the collection of neural signals via electrodes placed on the surface of a participant’s scalp.

For my research, I used an EEG headset called the EMOTIV EPOC X. This headset has 14 electrodes that make contact with a user’s scalp in order to collect the electrical activity of their brain. Once I did preliminary research on what has been done before and what is possible with technology like this, I began by exploring the software and looking at what EEG signals look like on a screen. I started by testing the EEG headset on myself so that I could gain more experience with the system and better design my experiment.

The purpose of my research was to create a predictive algorithm that takes an input of neural signals and outputs whether the user would like to open or close their fist.

The first step to this research was to collect data. In order to do this, I had a participant sit comfortably in a chair with their arms on a table, staying motionless unless they are prompted to move. They were then told to physically close their fists, physically open their fists, imagine closing their fists, and imagine opening their fists within many randomized trials.

The data was collected in the EMOTIV Pro software and exported to MATLAB for processing. In MATLAB, I used a band-pass filter to extract the beta frequency band for the signals correlating to physical movement, and the mu frequency band for the imagined movement signals.

For my predictive algorithm I used machine learning. Specifically, I used an artificial neural network (ANN) where I split my data 80% towards training and 20% towards testing.

My results were that my algorithm had a 41.8% accuracy in predicting the physical opening and closing of one’s fists and a 58.6% accuracy in predicting the imaginative opening and closing of one’s fists. Therefore, there was a low accuracy in my algorithm’s ability to predict fist motions based on EEG data.

Future improvements to this study would be to use a more reliable collection device, randomize the data better, extract different features, test the use of other algorithms, and connect the algorithm to a prosthetic hand for live manipulation of the hand.

I would like to thank OSCAR, Dr. Siddhartha Sikdar, Dr. Holger Dannenberg, Dr. Karen Lee, Amal Nadel, Christopher Kim, Binal Brahmbhatt, and Shriniwas Patwardhan for their help in completing this research project.

Thank you