INTRODUCTION

Given the large population of patients with intractable epilepsy, the automatic detection of seizure onset has sparked great interest among researchers over the past 20 years. In addition to providing a vital seizure alert to the patient, caregiver or a therapeutic device, it significantly eases the task of reviewing and labeling seizure segments in a patient’s EEG, a time-intensive task routinely done by neurologists. Implanting a device that performs both detection and closedloop suppression is the ultimate goal. Today, the Responsive Neurostimulator (RNS) by NeuroPace provides an FDAapproved therapy option to reduce the seizure frequency. However, RNS is bulky, limited in number of channels, and only relies on simple hard thresholding with moderate seizure classification accuracy.

classification accuracy. The power and area constraints imposed by implantable devices do not allow the implementation of sophisticated on-chip classification algorithms. Indeed, even the simple arithmetic operations performed in conventional classification methods, such as SVMs [1] and k-nearest neighbor (KNN) algorithms [2] can become very costly with increasing number of recording channels and higher sampling rates. With only simple comparator stages as their building blocks, decision trees (DTs) are a preferable solution to reduce hardware design complexity. Despite all their advantages, decision trees are unfortunately very susceptible to overfitting in seizure detection, particularly due to the high dimensionality of the feature space. This necessitates a careful design.