Epilepsy is one of the most common neurological disorders. Approximately 1% of people in the world suffer from epilepsy, and 25% of epilepsy patients cannot be healed by today’s available treatments [1, 2]. If seizures cannot be well controlled, the patients experience major limitations in family, social, educational, and vocational activities.
1.1 Motivation
Recently, numerous alternative techniques have been proposed, such us vagus nerve and deep brain stimulation devices [1, 2]. Most of the devices utilize open-loop controller to suppress the seizure. However, an open-loop controller drives a stimulator continuously or intermittently that causes high power consumption and the likelihood of neuronal damage. In contract, a closed-loop device combines a stimulator and seizure detector. Recently, one closed-loop epilepsy control system developed by NeuroPace called Responsive Neurostimulator (RNS®) System is in U.S. FDA clinical trials [2]. A closed-loop device can increase stimulus efficacy and reduce tissue damage over the long term. A closed-loop seizure controller drives a stimulator when a closed-loop device detects the seizure [2-4]. Despite additional hardware, a closed-loop device can increase stimulus efficacy and reduce tissue damage over the long term. As a result, compared with open-loop devices, closed-loop devices is more effective and attractive. In general, a robust on-line seizure detection method, which can drive antiepileptic device to suppress the seizure as early as possible when a seizure happens, is required for the development of a closed-loop seizure controller.
Recently, the implementation of hardware prototypes and biomedical signal processors has been proposed [5-19]. Among these projects, wavelet analysis, spectral analysis, entropy analysis, and variance analysis are applied to detect seizure events. Some closed-loop seizure controllers utilized analog to extract seizure features so epileptic seizure detection accuracy was high [15, 18]. Some seizure detection algorithm relied on powerful processing platform keeping real-time seizure detection and high detection accuracy [9-11, 13, 17]. However, the average response time for seizure detection is either longer than 5 seconds or often not mentioned in these works. Moreover, most of studies often use the discontinuous electroencephalogram (EEG) signal fragments to validate seizure detection algorithm;
nevertheless, it is deficient in validating the robustness of detection algorithm. As a result, a portable wireless online closed-loop seizure controller in freely moving rats was proposed [20-23], which validated seizure detection algorithm by using continuous online EEG signals.
Furthermore, the detection delay is shorter than 1 second. To summarize, an open-loop seizure controller with periodic stimulation is inaccurate and inefficient as shown in Fig. 1.1 (a). A closed-loop seizure controller with responsive stimulation which is proposed by other groups as mentioned above is more accurate and efficient; however, the detection delay is longer than 5 seconds as shown in Fig. 1.1 (b). Fig. 1.1 (c) shows that a closed-loop seizure controller with responsive stimulation is proposed by our group. When a seizure occurs, the responsive stimulator starts to suppress the seizure, and the detection delay is shorter than 1 second.
0 1 2 3 4 5 6 7 8 9 10
(b) Responsive Stimulation (Other groups)
Digitized EEG signal
(c) Responsive Stimulation (Our group)
Digitized EEG signal
Time(sec)
normal seizure
Detection delay < 1s
Inaccurate stimulation
Detection delay > 5s Inaccurate stimulation
Stimulation pulse
Fig. 1.1 A diagram of (a) periodic stimulation,
(b) responsive stimulation (other groups), and (c) responsive stimulation (our group).
1.2 Study Objective
In our previous work, using implied approximate entropy and 64-point fast Fourier transform, which contained large portion of digital processing, increased detection rate. In order to decrease complex calculation and hardware area, a liner least squares (LLS) was classifier in this project. Furthermore, it was observed that most of false detections occurred in slow-wave sleep (SWS) state, so adaptive thresholds were utilized to switch the threshold of the LLS for decreasing false detection rate. However, adaptive thresholds were obtained by using exhaustive key search in training phase; as a result, this method wasted a lot of time on training phase. In addition, implementation based on 8051-like microcontroller [24]
consumed more than 117mW to perform the real-time seizure detection.
In this study, it is proposed that using the mean and standard deviation of EEG training data searches adaptive thresholds rapidly. The new parameter determination method is faster than our previous work, and it can attain same performance. Moreover, the seizure detection algorithm in previous work is implemented in a RISC-like processor to suppress seizures. The flexibility, simplicity, and fixed instruction format of RISC [25] is feasible implementation with high processing performance. Although more complicate hardware architecture is used to realize real-time seizure detection, the RISC-like processor does not run algorithm at full speed in processing biomedical signals. As a result, a slower clock rate is applied to reduce the power and energy consumption of the proposed system. The continuous EEG signals of four Long-Evens rats are applied to the proposed biomedical signal processor. The results show the embedded processor is robustly processing 24 hours long-term and uninterrupted EEG sequence. In the future, the development of proposed processor will integrate analog front-end and antiepileptic circuitries into system-on-a-chip design for neural prosthesis applications.
1.3 Thesis Organization
The content of this thesis is organized as follows. Chapter 1 introduces the motivation and objective of this work. Chapter 2 describes the preparation of animal models and recorded EEG training data. In Chapter 3, the system architecture of proposed biomedical signal processor is described. Chapter 4 presents the epileptic seizure detection algorithm and the proposed parameter determination method. Chapter 5 demonstrates the hardware and firmware implementation. In Chapter 6, the evaluation procedure and measurement results are presented. Finally the conclusion and future work are made in Chapter 7.