Comparison between epileptic seizure prediction and forecasting based on machine learning.

Area of Science:

• Neurology
• Biomedical Engineering
• Data Science

Background:

• Epilepsy affects 1% of the global population, with anti-epileptic drugs failing in approximately one-third of patients.
• Current seizure warning devices often rely on prediction algorithms, which may not be optimal for all individuals.
• A shift towards seizure forecasting offers a probabilistic alternative to deterministic prediction methods.

Purpose of the Study:

• To explore seizure forecasting methodologies.
• To compare the performance of seizure forecasting against seizure prediction.
• To evaluate the suitability of forecasting for epilepsy warning devices.

Main Methods:

• Development of patient-specific seizure prediction and forecasting algorithms.
• Utilized Electroencephalogram (EEG) data from 40 patients in the EPILEPSIAE database.
• Employed classifiers including Logistic Regression, Support Vector Machines (SVM) ensemble, and Shallow Neural Networks (SNN) ensemble.

Main Results:

• Seizure forecasting demonstrated a significant increase in seizure sensitivity compared to prediction, up to 146%.
• The number of patients showing improvement over chance increased by up to 300% with forecasting.
• Forecasting algorithms showed enhanced performance across different patient-specific models.

Conclusions:

• Seizure forecasting provides a more effective approach for seizure detection than traditional prediction.
• Probabilistic forecasting may lead to improved quality of life for epilepsy patients through better warning systems.
• Forecasting's continuous risk assessment is a key advantage over event-triggered prediction.