PONTINE NUCLEUS AUDIO STIMULI DETECTION & MODELING FOR BRAIN MACHINE INTERFACE REHABILITATION OF CONDITIONAL LEARNING Hanan Shteingart 1 , Aryeh Taub 2 , Hagit Messer 1 Fellow of the IEEE Tel-Aviv University School of Electrical Engineering 1 Tel-Aviv University Department of Psychology 2 ABSTRACT In order to establish a brain-machine-interface (BMI) system that rehabilitates damaged cerebellum function of discrete motor learning, the detection of conditional and unconditional stimuli (CS and US) onset times based on electro-physiology recordings analysis is necessary. These signals are relayed through brainstem areas called Pontine Nucleus (PN) and the Inferior Olive (IO) respectively. In this paper we focus on the model based analysis of the PN and compare the expected model performance with the observed one with real samples. We suggest a model of multi-unit (MU) activity as filtered inhomogeneous Poisson pulses of evoked activity contaminated by homogenous spontaneous activity and thermal noise (Filtered Poisson-Poisson-Gaussian model). By assigning the likelihood into the generalize log likelihood test (GLRT), we show that the best expected feature is energy detection. The model parameters were estimated based on the recorded peri-stimuli-time-histogram (PSTH) by chi-square goodness-of-fit minimization. Monte Carlo simulation showed that the thermal noise can be neglected in respect to the spontaneous activity and also predicted the order of the observed empiric detection performance in terms of detection probability and area under the receiver operation characteristic (ROC) curve (AUC). Index Terms— Brain Machine Interface, Neural Decoding, Brainstem, Classical Conditioning, SVM, Electrophysiology, Multi Unit, Poisson Model, GLRT, Detection, ROC, AUC, Simplex, Goodness-of-fit. 1. INTRODUCTION Brain machine interfaces (BMI) are systems that enable the interaction between a living brain tissue and an external man-made-machine. The endeavor of the field is clinical rehabilitation application [1] . There has been much progress in the field of BMI on the last decade [1] which is mainly aimed at two periphery nervous system applications: sensory rehabilitation (e.g. bionic eye research) and motor control rehabilitation (e.g. limb prosthesis [2]). However, our research [3] is novel as we aim towards rehabilitation of a learning function in the central nervous system. The ReNaChip project [3] aims to tackle conditioned motor learning of discrete motor response as it considered being simple relative to other cognitive tasks [4; 5]. Specifically, it focuses on eye-blink conditioning by an auditory tone with an animal model (rats). In order to rehabilitate a malfunction cerebellum circuit, first we need to peak up the signals from the brainstem areas of the inputs and deliver them into the biomimetic chip that will reproduce the natural missing motor response (see Fig. 1). Here we focus on the estimating the onset time of the CS and US from the electro-physiological data (Fig 1). In this paper we suggest a Filtered Poisson-Gaussian model for the measurements and derived an optimal detector to find the discrimination performance sensitivity to the model's parameters using Monte Carlo simulations. Then, we estimated the model parameters which fit the observed. This allowed us to compare performance of feature based [5] detection of the real data. Moreover we could estimate upper bounds on performance from the model and compare it to real data. In section 2 we present the proposed model while in section 3 we derive its upper bound for performance in the homogenous or short time non-homogenous cases together with the necessary performance for our application. Section 4 describes the results of the empiric detection performance on real data (using features based supervised classifier), and compare it to simulation of the model. In section 5, we summarize and discuss the results. 2. MODELING OF THE MEASURED SIGNALS In the ReNaChip project, the data is collected by a multiunit electrode, due to the requirement for robustness and reliable operation over a long period (which supports possible clinical use). Oppose to single unit electrodes that aim to capture activities of a single or few neuron, such electrodes collect data from several neurons and integrate over their activities. We suggest to model the multi unit response activity () rt as the sum of three components () () () () rt nt st et    , 21 978-1-4244-2710-9/09/ 25.00 c  2009 IEEE Authorized licensed use limited to: Hebrew University. Downloaded on November 10, 2009 at 14:56 from IEEE Xplore. Restrictions apply.