Machine Learning Algorithms for Detection of Noisy/Artifact-Corrupted Epochs of Visual Oddball Paradigm ERP Data

Abstract

Electroencephalography (EEG) is a non-invasive monitoring method that tracks and records the neural activities of the brain. The time-locked capture of the EEG to the external stimuli is known as Event-Related Potential (ERP) and it can help elucidate how the brain responds to the stimuli. In general, EEG is an uneven mixture of neural and non-neural sources of activities and these non-neural (non-EEG) signals produce artifacts in the EEG that can decrease the SNR in experiments and may lead to erroneous conclusions about the effects of experimental manipulation. Thus, it is very important to remove artifacts from the recorded EEG prior to analysis. The most common artifacts impacting ERPs are eye-blink, eye-movement, and body-movement. These artifacts-corrupted data can be removed by visual inspection or by computer-automated signal processing methods. While these methods are suitable for post-processing of collected ERP applications, they not well-suited for real-time processing of continuous ERP data. This project seeks to address the challenges associated with real-time identification of artifacts by introducing a machine learning model that can screen ERP, detect and reject artifact-corrupted data epochs prior to signal analysis. In addition to enabling real-time pre-processing of streaming ERP data, the DBScan machine-learning methods explored here can provide up to 90% accuracy in the identification of artifacts-mixed ERP epochs. As a result, the findings of this study will help to improve the signal quality of ERP trials and will enable ERP to be used as a biomarker in real-world applications where streaming EEG data collection and analysis are required.