Artifact Detection and Removal from Electroencephalography(EEG) Data

Abstract

Electroencephalography (EEG) is a technique that records the electrical activity of the brain by placing electrodes on the scalp. It is used for a variety of purposes, including brain-computer interfaces, disease diagnosis, and determining cognitive states. Yet EEG signals are susceptible to noise from many sources, such as muscle and eye movements, and motion of electrodes and cables. The noise that contaminates EEG data is also called artifacts. Artifacts make it difficult to interpret the underlying neural information and can interfere with proper working of BCI devices. Consequently, artifact removal is an important preprocessing step for EEG data. There are two main challenges in removing artifacts from EEG data. First challenge is the lack of a paired training corpus, i.e., we do not have clean version of the noisy EEG signal. We have a set of noisy signals and a set of clean signals. Another challenge is the lack of supervisory information about how artifacts are manifested in the EEG signal. In this work, we propose three methods for detection and removal of artifacts that address these challenges. First, we propose a novel system that uses a weak supervisory signal to remove artifacts from EEG data. We use information that some noise is occurring, but not what the source of the noise is or how it is manifested in the EEG signal. To remove artifacts, we decompose EEG data into independent components using ICA, and these components form bags that are labeled and classified by a multi-instance learning algorithm that can identify the noise components for removal to reconstruct a clean EEG signal. We use this model to identify artifacts in real-world EEG dataset. We also perform extensive hyperparameter optimization for the model with the goal of improving accuracy without increasing execution time. Although our ICA-based method was trained using a weak supervisory signal, it needs large number of samples to converge and has large memory requirements. We solve these problems by proposing an online, fully automated, end-to-end system for artifact removal trained using unpaired training corpora. We propose a new architecture called AsymmetricGAN that uses a generative adversarial network for artifact removal. Even though the asymmetricGAN approach is online and end-to-end, training it is unstable and needs extensive hyperparameter optimization. We combined the previous two approaches to create a more stable end-to-end deep neural network for artifact removal. We do this by using a deep neural network that separates artifacts and EEG data using the independence assumption. Like the asymmetricGAN, the proposed network could be used in an online and end-to-end fashion. This network does not use an generative adversarial network, making it easier to train. We evaluated these approaches on a synthetic and a real-world EEG dataset.