Multiresolution analysis of multichannel neural recordings in the context of signal detection, estimation, classification and noise suppression.

The development of multichannel microprobe fabrication technology for recording neural activity in the brain has recently achieved a significant milestone towards integrating microimplanted device technology with research and clinical applications in neurophysiology. Recent probe designs have been able to integrate large number of sites on a single probe to provide neuroscientists with tools to record from large populations of cells. Advances in probe design are always governed by the feasibility of the associated communication and signal processing technology. Surprisingly, existing signal processing techniques are considerably behind the overwhelming advances in probe fabrication technology. We envision the problem of optimizing the information transfer from the microdevice as three fold: Noise suppression, Signal detection, and Blind Source Separation. We demonstrate that all three goals can be achieved by merging multiresolution analysis theory with array processing theory into a novel unified framework. In the noise suppression context, we show that we can near-optimally suppress the additive correlated noise by introducing a spatio-temporal decorrelation mechanism using eigendecomposition of a discrete wavelet transform representation of the array data followed by universal thresholding, a unique property of the multiresolution analysis. In the detection context, when no apriori knowledge is given about the signal and/or the noise processes, we formulate a transform domain Generalized Likelihood Ratio Test in the array case that overcomes the problem of estimating unknown noise parameters. The Blind Source identification problem is approached within the same context using an inherent invariance property of the signal subspace across multiresolution levels that enables characterization of each neural source. Results demonstrate that this framework provides the basis for simple and practical implementation in the structure of today's biosensor array technology without compromising issues of bandwidth, detection and classification. We show that the framework is capable of achieving substantial improvement in detection performance in severe noise conditions, and robustness to source nonstationarities and nonisotropic properties of the unknown medium under no constraints on the array design or prior knowledge of the signal parameters.Ph.D.Applied SciencesBiological SciencesBiomedical engineeringElectrical engineeringNeurosciencesUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/130311/2/3042147.pd