On-Chip Epilepsy Detection: Where Machine Learning Meets Patient-Specific Healthcare

Abstract

Detecting epileptic seizure from EEG/iEEG is challenging due to each patient's onset pattern being very different from each other. To detect such onsets (in real time), the hardware must learn each patient's ictal pattern from the EEG/iEEG traces. This paper describes feature extraction and usage of Support Vector Machine (SVM)-based classifiers to achieve a patient-specific seizure detection System-on-Chip (SoC). Linear SVM (LSVM), Non-Linear SVM (NLSVM), and the Dual Detector Architecture (D(2)A)-SVM are described and compared. We show that when there are a limited number of training sets, the D(2)A-SVM classifier performs well while minimizing the hardware cost. The SoC is implemented and verified.