Reliable but Wrong? Automated Detection of Heart Rate Artifacts as a Source of Bias in Pediatric EEG Data

Automated processing pipelines are increasing in popularity for analysis of EEG data, however care must be taken when using automated algorithms are applied outside their training dataset. This study assessed whether ICLABEL, a widely used artifact detection algorithm trained primarily on adult EEG, generalizes to pediatric EEG, with a focus on cardiac artifact which is easily identified by trained researchers yet prone to confounding automated detection due to spectral overlap with common EEG outcomes. Resting EEG child data from a randomized controlled trial (n = 47) was analyzed via ICLABEL and compared to manual artifact detection. ICLABEL miscategorized cardiac artifacts 100% of the time, with 66 of 92 components displaying a heart rate waveform labeled as 0% likelihood of cardiac artifact. To assess downstream impact, data with manual artifact removal was compared to data retaining heart rate components, examining effects on theta and beta power and theta/beta ratio, frequencies that overlap with cardiac artifact. Retention of heart rate artifacts significantly elevated theta and beta power estimates and introduced a spurious treatment effect on eyes-closed beta power, underscoring the importance of proper artifact removal before interpreting study outcomes. We recommend supervised use of the algorithm when applied to pediatric EEG or data that diverges from neurotypical adult data.