Wavelet-Based Neural Pattern Analyzer for Behaviorally Significant Burst Pattern Recognition Seetharam Narasimhan, Miranda Cullins, 1 Hillel J. Chiel and Swarup Bhunia Case Western Reserve University, Cleveland, OH Email: {sxn124, mjc35, hjc, skb21}@case.edu Abstract— Closed-loop neural prosthesis systems rely on accurately recording neural data from multiple neurons and detecting behaviorally meaningful patterns before representing them in a highly compressed form for wireless transmission over a limited-bandwidth link. We present a novel wavelet- based approach for detecting spikes, grouping them as bursts and building a dynamic vocabulary of meaningful burst pat- terns. Simulation results on pre-recorded in vivo multi-channel extracellular neural data from the buccal ganglion of Aplysia demonstrate the feasibility of behavior recognition as well as data compression (> 500X) by the proposed approach. I. INTRODUCTION Closed-loop neural prosthesis systems typically involve analysis of multi-channel recordings from behaviorally rel- evant neurons for predicting intended behavior before un- dertaking any preventive/corrective/assistive actions by ap- propriate neural stimulation/drug delivery. There are several design challenges to be addressed before such a system be- comes a reality. When hundreds of electrodes are employed for parallel recording, the transmission ability of the teleme- try device (e.g. its bandwidth) becomes insufficient and power-hungry [1], [2]. Therefore, it is extremely important to use on-chip electronics for pre-processing and compressing the recorded data that can be transmitted wirelessly with very low power dissipation. For closed-loop neural control, the implantable system must also perform real-time analysis of spike patterns from multiple channels and recognize behaviorally relevant patterns in the neural signal. In order to use the recorded neural information containing action potentials (spikes in intracellular voltage) mixed with background noise (biological and electrical), we need to de-noise the signal and detect the spikes. Such a compu- tational task for online, real-time data analysis and com- pression requires both efficient signal-processing algorithms and special-purpose customized hardware to implement them on-chip, since conventional microprocessor or DSPs would dissipate too much power and are too large in size for an implantable device. In this work, we address the issue of dig- ital signal processing for on-chip burst pattern identification, burst-level representation of neural signal and subsequently, recognition of meaningful behavior from bursts of neural activity. By representing neural signals in terms of bursts, we can achieve large data compression, thus saving precious bandwidth and reducing power dissipation. 1 Dr. Chiel was supported by the NIH Grant no. NS047073. Fig. 1. Simultaneous recording from four channels in the buccal ganglia of Aplysia. The multi-channel neural data can be analyzed to differentiate between two types of swallowing behaviors in the animal [3]. The experimental animal in our case is an invertebrate marine mollusk, the sea slug (Aplysia californica). It has an experimentally attractive nervous system, particularly because its neuromuscular dynamics can be readily related to its overall behavior. Studies relating neural activity to mus- cular dynamics have resulted in identification of well-defined burst patterns that can be readily correlated with meaningful behavior in Aplysia. Fig. 1 shows in vivo neural and muscular recordings from intact, behaving Aplysia during two types of swallowing behavior [3]. We use multi-resolution wavelet analysis [4] of recorded signals to de-noise the data, detect and sort spikes and then determine the on/off timing of a burst. Next, a hierarchical vocabulary is built to maintain spikes from different neurons, develop higher-order symbols in terms of bursts and recognize behaviorally meaningful burst patterns across multiple channels from in vivo neural and muscular recordings. II. RELATED WORK Major tasks to be performed by the digital signal process- ing unit are de-noising, detection of spikes, spike sorting, detection of bursts, efficient representation of spike data (for data compression) and recognition of meaningful patterns (that relate to the overall behavior of an animal) from the multi-channel recorded signal. Previous efforts at spike detection have been aimed at using simple thresholding schemes, with or without an adaptive threshold [5]. The existing hardware solutions for data compression and spike detection are based on thresholding schemes that do not preserve spike shapes [6], [7]. Wavelet-based detection has 30th Annual International IEEE EMBS Conference Vancouver, British Columbia, Canada, August 20-24, 2008 978-1-4244-1815-2/08/$25.00 ©2008 IEEE. 38