EEG Image Transformer for Harmful Brain Activity Classification.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Timely detection of critical brain activities like seizures is vital in neurocritical care.
• Current diagnostic methods can be time-consuming and require expert interpretation.
• Harmful brain activities include seizures, lateralized periodic discharges (LPD), and generalized rhythmic delta activity (GRDA).

Purpose of the Study:

• To develop and evaluate an automated EEG classification framework, EEG-DeiT, for detecting harmful brain activities.
• To improve diagnostic speed and accuracy in neurocritical care settings.
• To enhance patient outcomes through early intervention.

Main Methods:

• Proposed an EEG-DeiT framework fusing raw EEG and spectrogram images.
• Employed extensive data preprocessing including log transformation and standardization.
• Utilized the DeiT-EEG architecture for enhanced learning on clinical datasets.

Main Results:

• Achieved 97.65% accuracy in classifying harmful brain activities.
• Outperformed traditional Vision Transformers (ViT) in accuracy.
• Demonstrated expert-level performance and 3.8x faster convergence compared to conventional transformers.

Conclusions:

• EEG-DeiT enables real-time, automated classification of critical EEG patterns.
• The framework improves diagnostic efficiency and accuracy, aiding early intervention.
• This technology has significant clinical relevance for enhancing patient care in neurocritical settings.