Classification of self-limited epilepsy with centrotemporal spikes by classical machine learning and deep learning...

Area of Science:

• Medical Informatics
• Computational Neuroscience
• Pediatric Neurology

Background:

• Electroencephalogram (EEG) is crucial for epilepsy diagnosis in clinical settings.
• Distinguishing self-limited epilepsy with centrotemporal spikes (SeLECTS) from other epilepsies (non-SeLECTS) using EEG is diagnostically challenging due to similar abnormal discharges.
• Accurate classification of SeLECTS is vital for appropriate patient management and treatment.

Purpose of the Study:

• To develop and evaluate machine learning and deep learning models for classifying SeLECTS versus non-SeLECTS epilepsy.
• To compare the performance of classical machine learning with feature extraction against deep learning approaches using EEG data.
• To identify the most effective computational method for improving the diagnostic accuracy of SeLECTS.

Main Methods:

• Collected clinical EEG data from 33 pediatric patients (3-11 years) diagnosed with SeLECTS or non-SeLECTS.
• Applied classical machine learning by extracting sharp wave features (upslope, downslope, width at half maximum) and using Random Forest (RF) and Extreme Random Forest (ERF) classifiers.
• Utilized deep learning by directly inputting EEG data into a deep residual network (ResNet) for classification.

Main Results:

• The Extreme Random Forest (ERF) classifier achieved 73.15% accuracy, with an F1-score of 0.72, AUC of 0.75, and AUPRC of 0.63.
• The deep learning ResNet model demonstrated superior performance with 90.49% accuracy, an F1-score of 0.90, AUC of 0.96, and AUPRC of 0.92.
• Feature extraction methods showed good reliability in identifying relevant EEG biological features for SeLECTS.

Conclusions:

• Deep learning methods, particularly ResNet, exhibit significant potential for accurate SeLECTS classification in EEG analysis.
• The high accuracy achieved by ResNet surpasses traditional machine learning approaches for this specific epilepsy subtype.
• Computational methods provide valuable tools to aid clinicians in the challenging diagnosis of SeLECTS.