Computer-Aided Intracranial EEG Signal Identification Method Based on a Multi-Branch Deep Learning Fusion Model and...

Area of Science:

• * Neurology
• * Biomedical Engineering
• * Artificial Intelligence

Background:

• * Accurate analysis of invasive inspection intracranial EEG (iEEG) data is crucial for surgical intervention and managing drug-refractory epilepsy.
• * Identifying epileptogenic zones from iEEG signals presents a significant challenge in clinical practice.
• * Existing methods often rely on manual feature extraction, which can be time-consuming and subjective.

Purpose of the Study:

• * To propose a multi-branch deep learning fusion model for accurate identification of epileptogenic signals from iEEG data.
• * To integrate classical signal processing techniques with deep learning for enhanced feature extraction and signal detection.
• * To validate the model's effectiveness on public and private clinical iEEG databases.

Main Methods:

• * A hybrid model combining a classical approach (feature extraction + Bi-LSTM-AM) and a deep learning approach (1D-CNN) for end-to-end feature extraction.
• * Integration of two branches to obtain deep fusion features for improved epileptogenic signal identification.
• * Resampling techniques used to address data imbalance between epileptogenic and non-epileptogenic samples for robust validation.

Main Results:

• * The model achieved high sensitivity (97.78%), accuracy (97.60%), and specificity (97.42%) on the Bern-Barcelona iEEG database, outperforming current state-of-the-art methods.
• * Demonstrated strong performance on a large-scale clinical stereo EEG database with average intra-subject accuracy of 92.53% and cross-subject accuracy of 88.03%.
• * The fusion model effectively identified epileptogenic signals, showcasing its clinical utility.

Conclusions:

• * The proposed multi-branch deep learning fusion model offers a robust and accurate method for identifying epileptogenic sources in iEEG data.
• * This approach has the potential to significantly aid researchers and clinicians in developing automated systems for epilepsy source localization.
• * The model's high performance on diverse datasets suggests its value in improving epilepsy diagnosis and treatment planning.