University of Southampton
Bit-serial artificial neural networks for epilepsy seizure detection
Abstract
dc:description.abstractFifty million of the world’s population are afflicted with epilepsy and 80% of these epileptic patients lives in developing countries. It is crucial to develop a low cost, power saving and reliable home-based seizure detection system for those disabled individuals who have insufficient access to seizure detection equipment.<br/>This research presents three contributions. The first demonstrates that simple bitserial architecture can be used when designing extremely low-power and low-cost neural network processors to detect epileptic seizures. The proposed design is tailored to be cost effective by employing variable bit precision to allow for compromise between the detection accuracy and the hardware cost.<br/>The second contribution highlights extensive studies of epileptic seizure detection by DPU arrays, using bit-serial neural networks (BSNN) where the control module consists of only simple finite state machines. It has been demonstrated that epilepsy detection through such low-cost and low-energy dedicated neural network is feasible and there is potential for massively parallel network configuration. Different network configurations with variable numbers of network nodes and layers were designed and tested on FPGAs. The best performing version of the complete system has been implemented on an ALTERA Cyclone V FPGA which uses 3931 ALMs with an average recognition rate of 89%.<br/>The third contribution illustrates the development of a dedicated feature extraction component to be used as part of the proposed epilepsy detection system. Two different dedicated feature extraction hardware systems have been designed to provide inputs to the neural network in order to facilitate the classification of EEG waveforms. The EEG features extracted in this research are the slope and mean energy in EEG waveforms. Through multiple experiments, it was found that using a combination of both features as input to the proposed BSNN provides a detection accuracy of 90%.<br/>Results of this research have been published in three conference papers and also in the IEEE Journal on Translational Engineering in Health and Medicine.<br/>
Degree
thesis:*- Name dc:type.qualificationname
- Ph.D.
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- University of Southampton
- Year dc:date.issued
- 2021
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kueh, Si Mon
- Advisor dc:contributor.advisor
-
- Kazmierski, Tomasz
Rights
- Language dc:language
- English, eng