EEG-based harmful brain activity classification using deep learning and feature fusion.

Area of Science:

• Neuroscience
• Medical Technology
• Artificial Intelligence

Background:

• Harmful brain activity detection is critical for critically ill patients, yet specialist shortages increase mortality.
• Standardization of electroencephalography (EEG) terminologies has spurred research in automated brain activity analysis.
• An automated system is essential for timely detection and classification of harmful brain activities, enhancing patient safety.

Purpose of the Study:

• To develop and validate a novel pipeline for classifying harmful brain activities using EEG data.
• To implement a feature-level fusion scheme via a dual 1D convolutional neural network (CNN) for improved classification accuracy.
• To demonstrate the robustness and effectiveness of the proposed automated system.

Main Methods:

• A dual 1D convolutional neural network (CNN) model was employed for feature fusion.
• The Harvard Medical School (HMS) dataset was utilized for experimental validation.
• Ablation analysis and explainable AI techniques were applied to assess model robustness.

Main Results:

• The proposed pipeline achieved high accuracy in classifying harmful brain activities.
• The best performing model (model 8) demonstrated an accuracy of 98.14% with a loss of 0.05.
• 10-fold cross-validation confirmed the model's robustness, yielding an accuracy of 99%.

Conclusions:

• The developed automated pipeline effectively classifies harmful brain activities, including seizures and seizure-like patterns.
• The feature-level fusion approach using a dual 1D CNN shows significant promise for improving patient safety in critical care settings.
• This research contributes a robust and accurate automated system for EEG analysis, potentially reducing mortality rates.