Deep learning-based classification and segmentation of interictal epileptiform discharges using multichannel...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epilepsy diagnosis relies on identifying interictal epileptiform discharges (IEDs) in electroencephalographic (EEG) recordings.
• Accurate and efficient detection of IEDs is crucial for effective patient management.
• Current methods may struggle with the complexity and spatiotemporal patterns in multichannel EEG data.

Purpose of the Study:

• To develop a deep learning framework for high-accuracy classification and segmentation of IEDs in multichannel EEG.
• To preserve spatial information and interchannel interactions for improved IED detection.
• To create an AI-based toolbox for potential bedside epilepsy diagnosis.

Main Methods:

• Proposed U-IEDNet, a novel deep learning framework utilizing convolutional layers and bidirectional gated recurrent units for temporal feature extraction.
• Employed transformer networks as a spatial encoder to fuse multichannel EEG features and capture interchannel interactions.
• Implemented a U-shaped architecture with transposed convolutional layers for temporal decoding and IED segmentation, validated on public and private datasets.

Main Results:

• U-IEDNet achieved high performance metrics, including recall, precision, and F1-score, on both public and private EEG datasets.
• Demonstrated superior performance compared to existing methods in IED detection and segmentation.
• Analysis of attention weights enhanced model interpretability by elucidating spatial feature fusion based on brain network theory.

Conclusions:

• U-IEDNet shows significant potential for improving the accuracy and efficiency of IED detection in multichannel EEG.
• The developed AI toolbox may facilitate real-time epilepsy diagnosis at the bedside.
• This framework offers a promising advancement in leveraging artificial intelligence for neurological disorder diagnostics.