Evaluating machine- and deep learning approaches for artifact detection in infant EEG: classifier performance,...

Area of Science:

• Neuroscience and Biomedical Engineering
• Developmental Psychology

Background:

• Electroencephalography (EEG) is crucial for infant brain activity research but is prone to artifacts from infant movement and variability.
• Manual artifact detection in EEG is time-consuming and subjective, necessitating automated solutions for reliable analysis.

Purpose of the Study:

• To evaluate the performance of machine learning classifiers (Random Forest, SVM, DL) for automated artifact detection in infant EEG.
• To assess classifiers without requiring prior feature extraction, directly processing raw, filtered infant EEG data.

Main Methods:

• Collected EEG data from 294 infants (mean age 8.34 months) as part of the BRISE study.
• Analyzed 66,851 epochs, with 45% manually annotated as artifacts by an expert.
• Trained and tested Random Forest, SVM, and deep learning models on filtered EEG data.

Main Results:

• Random Forest (RF) and deep learning (DL) models achieved high balanced accuracy (0.873 and 0.881, respectively), significantly outperforming SVM (0.756).
• RF excelled with smaller datasets, while DL required larger datasets for optimal performance.
• A trade-off exists between classifier certainty, accuracy, and the proportion of data classified.

Conclusions:

• RF and DL classifiers offer effective automated artifact detection for infant EEG, reducing preprocessing time and enhancing study consistency.
• Automated artifact detection can facilitate large-scale infant EEG research and improve reproducibility through standardized preprocessing.