Resource efficient on-node spike sorting

Abstract

Current implantable brain-machine interfaces are recording multi-neuron activity by utilising multi-channel, multi-electrode micro-electrodes. With the rapid increase in recording capability has come more stringent constraints on implantable system power consumption and size. This is even more so with the increasing demand for wireless systems to increase the number of channels being monitored whilst overcoming the communication bottleneck (in transmitting raw data) via transcutaneous bio-telemetries. For systems observing unit activity, real-time spike sorting within an implantable device offers a unique solution to this problem. However, achieving such data compression prior to transmission via an on-node spike sorting system has several challenges. The inherent complexity of the spike sorting problem arising from various factors (such as signal variability, local field potentials, background and multi-unit activity) have required computationally intensive algorithms (e.g. PCA, wavelet transform, superparamagnetic clustering). Hence spike sorting systems have traditionally been implemented off-line, usually run on work-stations. Owing to their complexity and not-so-well scalability, these algorithms cannot be simply transformed into a resource efficient hardware. On the contrary, although there have been several attempts in implantable hardware, an implementation to match comparable accuracy to off-line within the required power and area requirements for future BMIs have yet to be proposed. Within this context, this research aims to fill in the gaps in the design towards a resource efficient implantable real-time spike sorter which achieves performance comparable to off-line methods. The research covered in this thesis target: 1) Identifying and quantifying the trade-offs on subsequent signal processing performance and hardware resource utilisation of the parameters associated with analogue-front-end. Following the development of a behavioural model of the analogue-front-end and an optimisation tool, the sensitivity of the spike sorting accuracy to different front-end parameters are quantified. 2) Identifying and quantifying the trade-offs associated with a two-stage hybrid solution to realising real-time on-node spike sorting. Initial part of the work focuses from the perspective of template matching only, while the second part of the work considers these parameters from the point of whole system including detection, sorting, and off-line training (template building). A set of minimum requirements are established which ensure robust, accurate and resource efficient operation. 3) Developing new feature extraction and spike sorting algorithms towards highly scalable systems. Based on waveform dynamics of the observed action potentials, a derivative based feature extraction and a spike sorting algorithm are proposed. These are compared with most commonly used methods of spike sorting under varying noise levels using realistic datasets to confirm their merits. The latter is implemented and demonstrated in real-time through an MCU based platform.