Analysis and Design of Low Energy Biomedical Signal Processor with Machine-Learning for Healthcare Application

Abstract

Epilepsy is a chronic, long-term central nervous system disorder that predisposes individual patients to recurrent epileptic seizure. Epileptic seizure is a sudden, brief episode of abnormal excessive or synchronous brain electrical activity that produces disruptive symptoms. These symptoms range from a lapse in attention, biting tongues to a body convulsion or even synopses, depending on the part of brain involved and patient’s age. Many designs have been introduced for seizure onset detection. However, due to some reasons such as the unacceptable false-positive detection rate, lack of seizure termination detection, epilepsy centers still do not use these designs in clinical routine. In this thesis, we introduce a novel patient-specific seizure onset/termination detection algorithm that can overcome some issues in existing designs, including 1) high false alarm rate, 2) lack of seizure termination detection, and 3) hardware-costly learning process. The analysis and verification with 14 patients from CHB-MIT database based on Matlab modeling is discussed in detail. The result shows that proposed algorithm can improve the detection rate to >93.5% for linear Support Vector Machine while suppressing the falsepositive detection rate to <2%, improve temporal resolution from 2 seconds to 1 second, and reduce detection latency. We also provide a comprehensive comparison with the state-of-art works, highlighting our advantages. In the end, we describe the implementation of 16-channel digital back-end based on 0.18um 1P6M technology, and highlight our novelties in design efficiency.

Date of Award	2014
Original language	American English
Supervisor	Jerald Yoo (Supervisor)