VLSI Architecture of NEO Spike Detection with Noise Shaping Filter and Feature Extraction Using Informative Samples Linh Hoang, Zhi Yang, Wentai Liu School of Engineering, University of California at Santa Cruz, CA 95064 {linh, yangzhi, wentai}@soe.ucsc.edu Abstract— An emerging class of multi-channel neural recording systems aims to simultaneously monitor the activity of many neurons by miniaturizing and increasing the number of recording channels. Vast volume of data from the recording systems, however, presents a challenge for processing and transmitting wirelessly. An on-chip neural signal processor is needed for filtering uninterested recording samples and performing spike sorting. This paper presents a VLSI architecture of a neural signal processor that can reliably detect spike via a nonlinear energy operator, enhance spike signal over noise ratio by a noise shaping filter, and select meaningful recording samples for clustering by using informative samples. The architecture is implemented in 90-nm CMOS process, occupies 0.2 mm 2 , and consumes 0.5 mW of power. I. I NTRODUCTION Recent advancements in neural signals recording sys- tems [1] enable neuroscientists and clinicians capture simultaneous activity of many neurons in the brain for analysis and studies. By using implantable mi- croelectromechanical systems (MEMS) multielectrode arrays [2] placed in the cerebral cortex, neuroscientists able to observe neurons communicate with one another by way of electrical activity, which is known as action potentials or simply as spikes. The direct applications for these multi-channel neu- ral recording and processing capable systems are the enabling technologies for neuroprosthetic devices— devices those can be controlled by thoughts. As reported in literature, neural signals recorded from monkey’s motor cortex were analyzed to build a relationship between neural activities and intended limbs movements then used to control a cursor on a computer screen or a robotic arm [3]. The positive achievements of the emerging technologies in brain-machine interface yearn for feedback mechanisms enabling the brain perceive in- formation through prosthetic sensors. Building a realistic bionic arm [4] is an example research that incorporates This work was supported in part by UCOP. The authors wish to acknowledge the support of TMSC for chip fabrication and ARM for physical IP. multi-channel neural recording and stimulation technolo- gies for actuating a robotic arm and perceiving senses from the prosthetic sensors respectively. To surmount the challenging requirements of an implantable neuroprosthetic device that is low power, small footprint, high performance signal processing and limited wireless data rate great efforts are aimed at developing hardware efficient algorithms and architec- tures. On-chip signal processing can reduce the wireless data transmission, provide real-time computing solu- tions to complex spike sorting problem and enable a closed-loop neuroprosthetic framework. In this paper, we will briefly present our spike detection with noise shaping filter and feature extraction algorithms using informative samples along with a detail description of a cost-effective hardware architecture. Section II reviews our new spike sorting algorithm through each of the processing steps. Section III describes the architecture and hardware implementation. The results and future works are presented in Section IV. II. ALGORITHMS Spike sorting is a process of assigning spikes to different neurons. The process can be broken down into three major steps as follow: spike detection, feature extraction and clustering. In this section, we present our spike detection and feature extraction algorithms those are implemented in hardware as describe in Section III. For the next step in spike sorting process, we briefly discuss our new clustering algorithm that uses results from feature extraction for grouping neurons. A. Spike Detection The purpose of spike detection is to identify a neural spike from ambient noise or idle period of a neuron. Unfortunately, signal-to-noise ratio (SNR) can be as low as 0dB making it difficult to detect accurately with a simple amplitude thresholding. A solution for this 0dB SNR spike detection is to employ a nonlinear energy operator filter [5]. 978 31st Annual International Conference of the IEEE EMBS Minneapolis, Minnesota, USA, September 2-6, 2009 978-1-4244-3296-7/09/$25.00 ©2009 IEEE