Utilization of temporal autoencoder for semi-supervised intracranial EEG clustering and classification

Area of Science:

• Neuroscience and Biomedical Engineering
• Machine Learning Applications in Healthcare

Background:

• Manual electroencephalography (EEG) analysis is time-consuming, subjective, and requires specialized expertise.
• The increasing volume of multi-channel EEG data from advanced technologies exacerbates analysis challenges.
• Supervised deep learning for EEG analysis is hindered by the scarcity of expert-annotated gold-standard data.

Purpose of the Study:

• To develop and validate a semi-supervised machine learning technique for EEG analysis.
• To reduce the reliance on extensive expert-provided annotations in clinical EEG research.
• To assess the method's efficacy in classifying EEG data and detecting specific neurological events.

Main Methods:

• A semi-supervised learning approach utilizing deep learning with minimal gold-standard labels.
• Implementation of a temporal autoencoder for dimensionality reduction of EEG data.
• Kernel density estimating (KDE) maps were employed using a small set of expert labels.

Main Results:

• The method achieved high performance in classifying intracranial EEG (iEEG) data (Pathologic, Normal, Artifacts) with AUROC scores up to 0.879.
• Acceptable classification results were obtained using only 100 gold-standard samples per category.
• The technique demonstrated generalization to novel patients for detecting Interictal Epileptiform Discharges (IEDs), achieving an AUROC of 0.877.

Conclusions:

• Semi-supervised machine learning offers a viable solution for analyzing large-scale EEG datasets with reduced annotation burden.
• The proposed temporal autoencoder-based method is effective and generalizable for clinical EEG analysis.
• This approach significantly lowers the barrier for applying deep learning in electrophysiology research and practice.