Towards Robust Seizure Type Classification via Curriculum Learning Paradigms.

Area of Science:

• Neurology
• Machine Learning
• Computational Neuroscience

Background:

• Seizure type classification (STC) is crucial for epilepsy diagnosis but challenging due to diverse seizure manifestations.
• Traditional machine learning (ML) models, especially deep networks (DNs), often suffer from overfitting and poor generalization due to random mini-batch training.
• The computational demands of deep networks like convolutional neural networks (CNNs) can be substantial.

Purpose of the Study:

• To develop a computationally efficient and generalizable ML framework for automated STC.
• To address the limitations of traditional training methods in DNs for STC.
• To improve the precision and accuracy of seizure classification using a novel approach.

Main Methods:

• Implemented a curriculum learning (CL) framework for STC, incrementally increasing task difficulty.
• Utilized the Temple University Hospital (TUH) dataset, dividing it into difficulty levels for staged training.
• Scheduled binary, ternary, and multiclass classifications for easy, medium, and hard tasks, respectively, within the CNN model.

Main Results:

• The CL-trained CNN achieved improved performance: 84.94% precision, 80.29% recall, 82.33% F1-score, and 80.29% accuracy.
• Demonstrated performance improvements of 2.02% (precision), 2.65% (recall), 2.58% (F1-score), and 2.65% (accuracy) over traditional training methods.
• Reduced training time by 170.88 seconds through the implementation of CL.

Conclusions:

• Curriculum learning (CL) effectively enhances the generalization and performance of deep networks for seizure type classification (STC).
• The proposed CL framework offers a computationally efficient solution for STC, improving upon traditional training methods.
• This approach holds significant clinical relevance for precise epilepsy diagnosis and improved patient outcomes.