Concept-drifts adaptation for machine learning EEG epilepsy seizure prediction

Area of Science:

• Neurology
• Machine Learning
• Biomedical Engineering

Background:

• Seizure prediction is vital for patient quality of life, yet 30% of patients are unresponsive to current treatments.
• Identifying the preictal state is key, but Electroencephalogram (EEG) data's dynamic nature (concept drift) necessitates adaptive prediction methods.
• Existing EEG-based seizure prediction methods often lack clinical applicability.

Purpose of the Study:

• To evaluate automatic concept drift adaptation techniques for enhancing seizure prediction accuracy.
• To compare three patient-specific seizure prediction algorithms against a control approach using EEG data.

Main Methods:

• Investigated three adaptive algorithms: Backwards-Landmark Window (SVM), Seizure-batch Regression (logistic regression), and Dynamic Weighted Ensemble.
• All methods used univariate linear features, Support Vector Machines (SVM) classifiers, retraining post-seizure, and Firing Power for alarm generation.
• Compared adaptive methods against a standard machine learning pipeline on 37 Temporal Lobe Epilepsy patients from the EPILEPSIAE database.

Main Results:

• The Backwards-Landmark Window approach achieved 0.75 ± 0.33 sensitivity and 1.03 ± 1.00 false positive rate per hour.
• This adaptive strategy successfully predicted seizures above chance for 89% of patients.
• The control approach only validated seizure prediction for 46% of patients.

Conclusions:

• Automatic concept drift adaptation, particularly the Backwards-Landmark Window method, significantly improves seizure prediction performance.
• These adaptive techniques offer a more robust and clinically relevant approach to seizure prediction in epilepsy.
• The findings highlight the importance of addressing dynamic data changes in EEG for effective seizure forecasting.