ObjectiveThis study aimed to develop and validate a system of specialized deep lightweight convolutional neural networks (CNN) to accurately detect specific artifact classes and demonstrate their advantage over traditional rule-based methods.
MethodsThree distinct CNN systems were trained on the Temple University Hospital EEG Artifact Corpus to identify eye movement, muscle-related and non-physiological artifacts, with each system optimized for an ideal temporal window size. The performance of the proposed system was compared with standard rule-based clinical detection methods in a held-out test set.
ResultsThe CNN systems significantly outperformed rule-based methods, with F1-score improvements ranging from +11.2% to +44.9%. Importantly, the results revealed distinct optimal temporal window lengths for each artifact type: 20s for eye movements (ROC AUC 0.975%), 5s for muscle activity (Accuracy 93.2%), and 1s for non-physiological artifact(F1-score 77.4%).
ConclusionThe results show that specialized, artifact-specific CNNs provide a more consistent and accurate solution for automated EEG artifact detection than traditional rule-based approaches
SignificanceThis study establishes a new benchmark for automated EEG quality control by validating one of the first open-source, specialized CNN systems for three distinct artifact classes, both high sensitivity and specificity.