Two-stage approach based on combination of one-class SVM and CNN for epileptic seizure identification

Many automated EEG seizure detection methods are able to perform well in terms of precision but fail to keep reasonable recall producing a lot of false positives. To solve this, we propose a two-stage framework combining classical outlier detection methods and neural networks, utilizing real-world EEG data. The first stage uses a one-class support vector machine to filter non-seizure activity, followed by a convolutional neural network to distinguish between false positives and true seizures. Evaluation of metrics tailored for epilepsy detection showed improved precision, with a slight decrease in recall for the proposed approach, suggesting that it may improve the performance of clinical decision support systems.