Optimal training dataset composition for SVM-based, age-independent, automated epileptic seizure detection

Area of Science:

• Biomedical Engineering
• Neurology
• Machine Learning

Background:

• Automated seizure detection aids health professionals in minimizing brain damage through timely treatment.
• Current research primarily focuses on EEG feature computation and classification, with limited investigation into optimal training dataset composition for patient-independent seizure detection.

Purpose of the Study:

• To evaluate the performance of classifiers trained on different datasets for automated, age-independent seizure detection.
• To determine the optimal dataset composition for training a seizure detection classifier that performs well across different age groups.

Main Methods:

• Trained a support vector machine (SVM) classifier on three datasets: neonatal EEG, adult EEG, and combined neonatal and adult EEG.
• Evaluated two versions of each classifier: one with feature normalization and one without.
• Used the area under the receiver operating characteristics curve (AUC) as the performance measure, testing on both neonatal and adult EEG datasets.

Main Results:

• Classifiers trained on combined neonatal and adult EEG data achieved high performance (AUC 0.90 for neonatal, 0.93 for adult seizure detection).
• Neonatal seizure detection performance was significantly worse when trained solely on adult EEG data.
• Optimal adult seizure detection was achieved with classifiers trained on adult EEG or combined EEG data.

Conclusions:

• Age-independent seizure detection is feasible by training a single classifier on a diverse dataset including both neonatal and adult EEG data.
• Including EEG data from all relevant age categories in the training set is crucial for accurate age-independent seizure detection, particularly for neonatal seizure detection.
• The composition of the training dataset significantly impacts the accuracy of automated, age-independent seizure detection.