Multi-Head Self-Attention Model for Classification of Temporal Lobe Epilepsy Subtypes

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Temporal Lobe Epilepsy (TLE) is a chronic neurological disorder affecting over 70% of drug-resistant epilepsy patients globally.
• Accurate diagnosis of TLE and its subtypes is challenging due to complex etiology and reliance on clinician expertise, with a lack of specific biomarkers.
• Understanding TLE's brain network is crucial for diagnosis and treatment development.

Purpose of the Study:

• To develop and evaluate a novel multi-head self-attention model (MSAM) for enhanced classification of TLE subtypes.
• To investigate the potential of integrating self-attention mechanisms and multilayer perceptrons for improving TLE diagnosis.
• To compare the performance of MSAM against established machine learning algorithms.

Main Methods:

• A multi-head self-attention model (MSAM) was proposed, combining self-attention and multilayer perceptron.
• The MSAM was trained and tested on a Magnetoencephalography (MEG) dataset collected from TLE patients.
• Performance was evaluated against Convolutional Neural Network (CNN), Support Vector Machine (SVM), and Random Forest (RF) models.

Main Results:

• The MSAM achieved superior performance, with an accuracy of 83.6%, recall of 90.9%, precision of 90.7%, and F1-score of 83.4%.
• The study assessed the effectiveness of different numbers of attention heads to optimize the model.
• Ablation tests confirmed the robustness, effectiveness, and generalizability of the MSAM for TLE subtype classification.

Conclusions:

• The proposed MSAM demonstrates significant potential as a tool for accurate TLE subtype classification.
• The self-attention mechanism effectively learns signal location weights, enhancing classification accuracy.
• MSAM offers a promising, data-driven approach for improving the diagnosis of Temporal Lobe Epilepsy.