On Brain Computer Interfaces

Some ten years ago we worked on a project involving Brain-computer interfaces. We had just finished reading the book The Future of the Mind by Michio Kaku, which got us interested in the field when the opportunity came along to work on a related project, so we took it. This remains our only work in that field, as we went on to work on other projects and we therefore don't claim to be experts. Recently we stumbled upon our old code and paper back from 2015, which made us want to blog about the project and reflect on our work. 

EEG cap used as Brain-computer Interface by midjourney

First we want to describe the setup of our experiment and the hypothesis we were investigating with it. The interface we used was an EEG cap, measuring electrical activities coming from the brain of our test subjects. We had several test subjects perform a task that involves moving a cursor once they see an event trigger on the screen in front of them. We added a measuring device to the screen, so we'd know exactly when the trigger was shown. The task they'd perform was timed and the faster they completed, the better their score. Now, the question we were researching was, is there a measurable mental state, that can be used as an indicator for the upcoming performance? I.e. is it possible to determine if an upcoming task is going to be performed well, by looking at test subjects brain activity, right before the event is triggered?

Screenshot of the task that test subjects had to perform. The box on the left had to be moved into the larger box on the right using a curser. The color of the left box changes to indicate that the task has begun. If test subjects moved the cursor too early the task counted as failed.

Our project was done in multiple stages, at first we were coding up the experiment, next we spent some weeks performing it on multiple people and finally we got to data analysis. The analysis was done in three stages as well, data cleaning and preparation, e.g. taking the EEG time series and identifying the relevant spots right before the triggered events. This involved a lot of Fourier transformation. Next we trained ML models, using EEG data as features, trying to predict performance. And lastly we evaluated our results (including sanity checks, cross-validation etc.) and made them into a paper. On average we were able to predict performance with an 68% accuracy.

Figure from out paper, that shows signals for different sensors of the EEG cap. Each sensor has a name which indicates its location on the head of the test subject. The signals shown here are phase similarities positioned on the event trigger at zero. The signals are filtered to 8-13 Hz and we had shown it to discuss the quality of the data and its preprocessing.

This concludes our work in the field of BCI. While we don't expect to ever work in the field again, we do enjoy observing what is happening, e.g. Neural link announced last year, to test if their device is able to control a robotic arm. We also used to have some industry contact, to e.g. brainproducts and we plan to include at least one BCI related book into our reading agenda for this year.