A 16-channel, 1-second latency patient-specific seizure onset and termination detection processor with dual detector architecture and digital hysteresis

Abstract

This paper presents an area-power-efficient 16-channel seizure onset and termination detection processor with patient-specific machine learning techniques. This is the first work in literature to report an on-chip classification to detect both start and end of seizure event simultaneously with high accuracy. Frequency-Time Division Multiplexing (FTDM) filter architecture and Dual-Detector Architecture (D2A) is proposed, implemented and verified. The D2A incorporates two area-efficient Linear Support Vector Machine (LSVM) classifiers along with digital hysteresis to achieve a high sensitivity and specificity of 95.7% and 98%, respectively, using CHB-MIT EEG database [1], with a small latency of 1s. The overall energy efficiency is measured as 1.85μJ/Classification at 16-channel mode.