Retraining and evaluation of machine learning and deep learning models for seizure classification from EEG data

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Electroencephalography (EEG) is crucial for epilepsy diagnosis.
• Manual seizure annotation in EEG data is time-consuming.
• Machine learning models aim to automate seizure detection but face accuracy gaps in clinical practice.

Purpose of the Study:

• To reproduce and assess the accuracy of various machine learning models for automated seizure detection in EEG.
• To benchmark model performance across public datasets and local patient data.
• To identify effective models and contribute data for clinical integration.

Main Methods:

• Reproduced and evaluated numerous machine learning models, including deep learning networks.
• Trained and initially tested models on three public EEG datasets.
• Further tested models on a manually annotated EEG dataset from a local patient.

Main Results:

• Random forest and convolutional neural networks performed best on public datasets.
• A significant accuracy reduction was observed when testing on local patient data, particularly for the neural network.
• Performance disparities indicate challenges in generalizing models to diverse clinical data.

Conclusions:

• Current machine learning models for EEG seizure detection require improvement for clinical application.
• Retrained models and newly available data may enhance the accuracy of AI tools in epilepsy diagnosis.
• Bridging the gap between research accuracy and clinical utility is essential for AI integration.