Concept-Drifts Adaptation For Machine Learning EEG Epilepsy Seizure Prediction

Abstract The administration of antiepileptic drugs or surgical interventions fails to control seizures in about 30% of patients. Seizure prediction is a viable strategy for enhancing their quality of life because it can be used in intervention or warning systems. These systems may disarm seizures or, at the very least, lessen their adverse effects. Identifying the preictal interval, which marks the change from regular brain activity to a seizure, is critical to this research field. Even though several predictive studies applied various Electroencephalogram based methodologies, only some have been used in medical devices, and none have been clinically applicable. Recent studies have shown that tracking and handling data changes with time, known as concept drifts is highly relevant in seizure prediction; therefore, it is essential to develop methods able to automatically detect and address changes in context without human intervention. In this work, we aimed to evaluate the impact of automatic concept drift adapting methods in seizure prediction. We tested approaches to predict seizures while adapting to concept drifts during the model’s learning process; for that, we proposed and compared to the Control three patient-specific seizure prediction approaches with a 10-minute seizure prediction horizon: a seizure prediction algorithm incorporating a window adjustment method by optimising performance with Support Vector Machines1 (Backwards-Landmark Window), a seizure prediction algorithm incorporating a data-batch (seizures)selection method using a logistic regression2 (Seizure-batch Regression), and a seizure prediction algorithm with a dynamicintegration of classifiers3 (Dynamic Weighted Ensemble). The proposed methodologies included a retraining process after eachseizure and combined a set of univariate linear features with classifiers based on Support Vector Machines. The Firing Power wasused as a post-processing technique to generate alarms before seizures. Considering a group of 37 patients with Temporal LobeEpilepsy from the EPILEPSIAE database, the best-performing approach (Backwards-Landmark Window) aimed to select datafrom the concept closest to the preictal period of the last training seizure; this led to results of 0.75 ± 0.33 for sensitivity and 1.03 ± 1.00 for false positive rate per hour. Even though the best-performing approach statistically validated 89% of the patients with the surrogate predictor, it is necessary to determine the maximum false positive rate appropriate for each intervention system.