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

Abstract

Abstract Objective. Automated detection of artefact in stimulus-evoked electroencephalographic (EEG) data recorded in neonates will improve the reproducibility and speed of analysis in clinical research compared with manual identification of artefact. Some studies use very short, single-channel epochs of EEG data with little recorded EEG per infant—for example because the clinical vulnerability of the infants limits access for recording. Current artefact-detection methods that perform well on adult data and resting-state and multi-channel data in infants are not suitable for this application. The aim of this study was to create and test an automated method of detecting artefact in single-channel 1500 ms epochs of infant EEG. Approach. A total of 410 epochs of EEG were used, collected from 160 infants of 28–43 weeks postmenstrual age. This dataset—which was balanced to include epochs of background activity and responses to visual, auditory, tactile and noxious stimuli—was presented to seven independent raters, who independently labelled the epochs according to whether or not they were able to visually identify artefacts. The data was split into a training set (340 epochs) and an independent test set (70 epochs). A random forest model was trained to identify epochs as either artefact or not artefact. Main results. This model performs well, achieving a balanced accuracy of 0.81, which is as good as manual review of data. Accuracy was not significantly related to the infant age or type of stimulus. Significance. This method provides an objective tool for automated artefact rejection for short epoch, single-channel EEG in neonates and could increase the utility of EEG in neonates in both the clinical and research setting.