EEG-Based Speech Decoding Using a Machine Learning Pipeline

Abstract: his project aims to find a method that will help fill the information gaps in electroencephalography (EEG) brain-computer interfaces (BCI) research, by creating a pipeline method that allows for quicker research iterations than current state-of-the-art methods. The pipeline method is a multi-step processstarting from the recording EEG data from a subject performing a thought paradigm action, continuing with processing and decoding of the data, and ending with visualization and analysis the decoded results. Thought paradigms are in this project defined as different ways that the subject can think, with different words and different ways of thinking of those words. The pipeline will utilize various machine learning methods to be able to reach the two main goals of quickly being able to analyze and compare different paradigms and methods. Regarding the accuracy of the models, a minimum level of higher than random chance accuracies is needed if the pipeline should be considered to be useful for analyzing and comparing paradigms and methods, while a higher level of having accuracies comparable with state-of-the-art methods will allow for comparisons with paradigms and methods from other research methods as well. In the pipeline, various simple feature extraction methods are tested, such as the Fourier transform (FT) and low pass filtering. As well as features based on covariance between channels and data gradients. A specific way to baseline correct the features is also proposed and tested. The results of the project show that the pipeline method is a viable way of quickly testing and comparing paradigms and methods. With results that are comparable to state of the art methods. While also allowing for quick iteration and comparison. Future possibilities using this method are discussed