Automated EEG Channel and Epoch Quality Control

Abstract

Introduction

Electroencephalography (EEG) is a neuroimaging method widely used in neuroscience, clinical applications, and research in general. Part of every pre-processing routine of EEG signals involves the annotation of bad channels and bad epochs. In EEG, a bad channel refers to an electrode or channel producing poor-quality or unreliable data due to factors such as a malfunctioning electrode or poor contact between the electrode and the scalp. A bad epoch refers to a specific segment of time in the EEG recording that is compromised and cannot be used for further analysis. This can occur due to a variety of factors like movement, blinks, horizontal and vertical eye movement, or other physiological and non-physiological artifacts that can distort or obscure the EEG signal. Bad epochs and channels are typically removed from EEG signals using manual methods that are labor-intensive and often impossible to reproduce or automated methods that usually rely on basic statistical approaches [1] that usually do not give optimal results.

We aim to develop and explore the reliability and accuracy of modern machine learning methods for their detection in EEG recordings. Our objective is to create a solution that is both reproducible and gives good-quality annotations of manual methods.

Method

We plan to develop several machine learning models for the automatic detection of bad epochs and bad channels that are compatible with 32-, 64- and 128-channel EEG caps from diverse groups of EEG experiments. The datasets that are currently available consist of 16 experiments averaging two dozen participants.

During the initial modeling phase, we will apply traditional machine learning methods such as  K-nearest neighbors (KNN) or Support Vector Machines (SVM). In the subsequent phase, we plan to develop more sophisticated models utilizing deep-learning approaches such as Recurrent Neural Networks (RNN).

To validate the results of the models, we will compare the performance of our models with most of the popular EEG pipelines (e.g. [2]) that are equipped with automatic rejection of bad epochs/ channels.

Limitations

A major limitation of this study is the diversity, quality, and quantity of the available data for machine learning. As machine learning models are only as effective as the data they are trained on, the performance and generalizability of the models may be affected by the limitations of the dataset.

References

[1] H. Nolan, R. Whelan, and R. B. Reilly, “Faster: Fully automated statistical Thresholding for EEG artifact rejection,” Journal of Neuroscience Methods, vol. 192, no. 1, pp. 152–162, 2010. doi:10.1016/j.jneumeth.2010.07.015

[2] N. Bigdely-Shamlo, T. Mullen, C. Kothe, K.-M. Su, and K. A. Robbins, “The Prep Pipeline: Standardized preprocessing for large-scale EEG analysis,” Frontiers in Neuroinformatics, vol. 9, 2015. doi:10.3389/fninf.2015.00016