Unsupervised neural spike identification for large-scale, high-density micro-electrode arrays

Abstract

This work deals with the development and evaluation of algorithms that extract sequences of single neuron action potentials from extracellular recordings of superimposed neural activity - a task commonly referred to as spike sorting. Large ($>10^3$ electrodes) and dense (subcellular spatial sampling) CMOS-based micro-electrode-arrays allow to record from hundreds of neurons simultaneously. State of the art algorithms for up to a few hundred sensors are not directly applicable to this type of data. Promising modern spike sorting algorithms that seek the statistically optimal solution or focus on real-time capabilities need to be initialized with a preceding sorting. Therefore, this work focused on unsupervised solutions, in order to learn the number of neurons and their spike trains with proper resolution of both temporally and spatiotemporally overlapping activity from the extracellular data alone. Chapter (1) informs about the nature of the data, a model based view and how this relates to spike sorting in order to understand the design decisions of this thesis. The main materials and methods chapter (2) bundles the infrastructural work that is independent of but mandatory for the development and evaluation of any spike sorting method. The main problem was split in two parts. Chapter (3) assesses the problem of analyzing data from thousands of densely integrated channels in a divide-and-conquer fashion. Making use of the spatial information of dense 2D arrays, regions of interest (ROIs) with boundaries adapted to the electrical image of single or multiple neurons were automatically constructed. All ROIs could then be processed in parallel. Within each region of interest the maximum number of neurons could be estimated from the local data matrix alone. An independent component analysis (ICA) based sorting was used to identify units within ROIs. This stage can be replaced by another suitable spike sorting algorithm to solve the local problem. Redundantly identified units across different ROIs were automatically fused into a global solution. The framework was evaluated on both real as well as simulated recordings with ground truth. For the latter it was shown that a major fraction of units could be extracted without any error. The high-dimensional data can be visualized after automatic sorting for convenient verification. Means of rapidly separating well from poorly isolated neurons were proposed and evaluated. Chapter (4) presents a more sophisticated algorithm that was developed to solve the local problem of densely arranged sensors. ICA assumes the data to be instantaneously mixed, thereby reducing spatial redundancy only and ignoring the temporal structure of extracellular data. The widely accepted generative model describes the intracellular spike trains to be convolved with their extracellular spatiotemporal kernels. To account for the latter it was assessed thoroughly whether convolutive ICA (cICA) could increase sorting performance over instantaneous ICA. The high computational complexity of cICA was dealt with by automatically identifying relevant subspaces that can be unmixed in parallel. Although convolutive ICA is suggested by the data model, the sorting results were dominated by the post-processing for realistic scenarios and did not outperform ICA based sorting. Potential alternatives are discussed thoroughly and bounded from above by a supervised sorting. This work provides a completely unsupervised spike sorting solution that enables the extraction of a major fraction of neurons with high accuracy and thereby helps to overcome current limitations of analyzing the high-dimensional datasets obtained from simultaneously imaging the extracellular activity from hundreds of neurons with thousands of electrodes.