TEICUN: A Transformer Encoder Infused Convolutional UNet for EEG Motion Artifact Removal

The World Stroke Organisation (WSO) states that 1 in 4 people over the age of 25 will have a stroke in their lifetime [1]. Intervention is required to help rehabilitate stroke survivors and help them regain the maximum quality of life. Current rehabilitation techniques focus physical exercise and assessment but often do not monitor the brain directly. This would be useful in the treating of brain injury to be able to give more direct feedback and inform decision making. Currently, methods of assessing brain activity are limited by cost and their ability to work in motion. EEG is one of the cheapest methods for reading brain activity but suffers from artifacts caused by motion. This research aimed to address this issue by removing motion contamination from EEG signals. To do this, existing processing pipelines were assessed. These assessments covered both traditional and machine learning algorithms with the aim of identifying areas for innovation. Using PyTorch, and other available libraries, this research evaluated existing preprocessing methods and proposed new methods with a new machine learning architecture. The resulting architecture is a convolutional UNet with bottleneck replaced with a transformer encoder. The proposed preprocessing steps achieved a significant temporal correlation of 86.883% and a 26.896 dB improvement in SNR. These improvements highlight the method’s ability to preserve signal integrity while removing motion artifacts. Notably, the network accomplished this performance with around 173,000 fewer parameters than alternative methods, demonstrating state-of-the-art performance. This reduction in parameters reduces computational costs making it more suitable for real-time application in resource-constrained environments such as a clinical setting. However, comparison with other techniques is difficult because of differences in preprocessing techniques. The nuances of this are also discussed in detail.