Interpretable EEG seizure prediction using a multiobjective evolutionary algorithm.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Drug-resistant epilepsy affects approximately one-third of epilepsy patients, highlighting the need for effective seizure prediction.
• Current seizure prediction models often use fixed pre-ictal periods and deep learning, leading to "black-box" systems that lack clinical trust.
• Interpretability and patient-specific approaches are crucial for advancing seizure prediction technology.

Purpose of the Study:

• To develop an evolutionary seizure prediction model that automatically identifies optimal features and patient-specific pre-ictal periods.
• To enhance model interpretability, providing insights into seizure generation and algorithmic decisions.
• To improve patient comfort and trust in seizure prediction systems for drug-resistant epilepsy.

Main Methods:

• An evolutionary algorithm was employed to automatically search for the best feature set and pre-ictal period.
• The model was designed to provide patient-specific, interpretable insights.
• The methodology was tested on 238 seizures from 93 patients using scalp recordings, comparing results against a seizure surrogate predictor and a standard machine learning pipeline.

Main Results:

• The evolutionary model achieved performance above chance for 32% of patients.
• A standard machine learning pipeline marginally outperformed the evolutionary model, validating 35% of patients.
• Out of 54 patients performing above chance, 38% were validated solely by the evolutionary method, and 44% solely by the control method.

Conclusions:

• The study highlights the need for diverse methodologies in seizure prediction, as different approaches validate different patient cohorts.
• Patient-specific, interpretable seizure prediction models may enhance clinical understanding and adoption.
• The findings suggest that a personalized approach to seizure prediction is essential for improving outcomes in drug-resistant epilepsy.