Calibrating Deep Learning Classifiers for Patient-Independent Electroencephalogram Seizure Forecasting

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Machine learning, particularly deep learning, shows promise for analyzing electroencephalogram (EEG) signals for seizure forecasting.
• Current methods often rely on randomized cross-validation, which is insufficient for clinical applications requiring patient-independent testing.

Purpose of the Study:

• To develop and evaluate a calibration pipeline for fine-tuning deep learning models for seizure forecasting on independent patients.
• To assess the effectiveness of this calibration on patient-independent EEG data.

Main Methods:

• A simple calibration pipeline was implemented to fine-tune deep learning models using minimal data from new patients.
• The procedure was evaluated on two large EEG datasets from epileptic subjects.
• Performance was compared between deep learning and traditional machine learning algorithms.

Main Results:

• The calibration procedure significantly improved seizure forecast accuracy by over 20% on average for independent patients.
• Performance gains were observed systematically across all independent patients tested.
• The method demonstrated effectiveness for both deep learning and feature-based machine learning models.

Conclusions:

• A patient-independent calibration method can enhance seizure forecasting accuracy, making deep learning models more viable for clinical use.
• Realistic validation methods are crucial for comparing seizure prediction algorithms and developing robust healthcare tools.
• The proposed calibration requires at least one seizure event per patient for model adaptation.