Hybrid machine learning method for a connectivity-based epilepsy diagnosis with resting-state EEG

Area of Science:

• Computational Neuroscience
• Medical Imaging Analysis
• Machine Learning in Medicine

Background:

• Epilepsy diagnosis often relies on lengthy and expensive electroencephalographic (EEG) monitoring.
• Graph metrics for brain connectivity offer novel diagnostic approaches.
• Deep learning can potentially learn connectivity patterns directly from EEG data.

Purpose of the Study:

• To evaluate the performance of a convolutional neural network (CNN) algorithm for epilepsy diagnosis using EEG data.
• To explore the potential of deep learning in identifying brain connectivity alterations associated with epilepsy.
• To develop a more efficient and accurate diagnostic method for epilepsy.

Main Methods:

• A CNN algorithm was applied to directed Granger causality (GC) connectivity measures derived from resting-state surface EEG recordings.
• Data included 30 subjects with epilepsy and 30 controls, with 50-second recordings.
• An ensemble of CNN models was trained on variously prepared data and electrode combinations.

Main Results:

• The trained CNN identified reduced delta band connectivity in frontal regions and increased left lateralized frontal-posterior gamma band connectivity.
• An ensemble of CNN models achieved an 85% diagnosis accuracy with an 85% F1 score.
• Differential connectivity patterns were identified as potential neuromarkers for epilepsy.

Conclusions:

• Appropriate preparation of connectivity data allows generic CNNs to detect discriminative epileptic features.
• The identified differential patterns may offer insights into cognitive alterations in epilepsy.
• This deep learning approach shows promise as a valuable clinical decision support system for epilepsy diagnosis.