Neural spike feature extraction and data classifications.

摘要

A thorough study on processing in vivo neural recording data is presented in the thesis, where topics include extracellular neural spike modeling, neural noise/interface noise/electronic noise modeling, noise reduction, spike detection, spike feature extraction, nonparametric clustering, and DSP implementation. Contributions are summarized as follows. First, through modeling neural spikes, neuronal geometry signatures for improving waveform differentiation are identified. Second, after modeling and measuring noise sources of in vivo neural interface, it reaches an important conclusion that the dominant noise sources are neurons themselves followed by electrode interface, pointing out that the major efforts in the literature to reduce electronics noise are inefficient. Third, instead of developing low noise recording instrumentation, two algorithms based on adaptive band-pass filtering and noise shaping to deal with the dominant neural noise are developed. These two algorithms are independent of spike sorting algorithms and improve the signal to noise ratio by 3-6 dB. Fourth, an informative sample based feature extraction algorithm has been developed to enable real-time extraction of features from neural spikes. When compared with conventional approaches such as PCA, wavelets, or max-min, it demonstrates dramatically improved speed and accuracy. Fifth, to allow unsupervised execution of partitioning neural spike feature space, a fast and robust clustering algorithm called evolving mean shift is established. Last, preliminary hardware implementations have been carried out [1, 8-13], enabling the algorithms to function as a small device with real-time low-latency processing and low-power operation. Such efforts to enable portable and implantable systems are important to allow the study of complex behavior in neuroscience experiments, closed loop deep brain stimulation, and cortical controlled neuromuscular prostheses.