A machine learning artefact detection method for single-channel infant event-related potential studies

Area of Science:

• Neuroscience
• Biomedical Engineering
• Clinical Research

Background:

• Automated artefact detection in neonatal electroencephalography (EEG) is crucial for faster, more reproducible clinical research.
• Existing methods are unsuitable for short, single-channel infant EEG epochs due to data limitations and infant vulnerability.
• Developing a specialized automated method is necessary for this specific application.

Purpose of the Study:

• To create and evaluate an automated method for detecting artefacts in 1500 ms, single-channel infant EEG epochs.
• To address the limitations of current artefact detection techniques in neonatal EEG analysis.
• To enhance the utility of EEG in neonatal clinical and research settings.

Main Methods:

• A dataset of 410 EEG epochs from 160 infants (28-43 weeks postmenstrual age) was utilized.
• Epochs included background activity and responses to various stimuli, labeled by seven independent raters.
• A random forest model was trained on 340 epochs and tested on 70 epochs for artefact detection.

Main Results:

• The random forest model achieved a balanced accuracy of 0.81, comparable to manual review.
• Accuracy was consistent across different infant ages and stimulus types.
• The model effectively identified artefacts in short, single-channel neonatal EEG data.

Conclusions:

• The developed automated method provides an objective tool for artefact rejection in neonatal EEG.
• This technique can significantly improve the efficiency and reliability of analyzing short-epoch, single-channel infant EEG data.
• The automated tool has the potential to increase the application and value of EEG in neonatal care and research.