Resource-efficient algorithms and circuits for highly-scalable BMI channel architectures

摘要

The study of the human brain has for long fascinated mankind. This organ that controls all cognitive processes and physical actions remains, to this day, among the least understood biological systems. Several billions of neurons form intricate interconnected networks communicating information through through complex electrochemical activities. Electrode arrays, such as for EEG, ECoG, and MEAs (microelectrode arrays), have enabled the observation of neural activity through recording of these electrical signals for both investigative and clinical applications. Although MEAs are widely considered the most invasive such method for recording, they do however provide highest resolution (both spatially and temporally). ududDue to close proximity, each microelectrode can pick up spiking activity from multiple neurons. This thesis focuses on the design and implementation of novel circuits and systems suitable for high channel count implantable neural interfaces. Implantability poses stringent requirements on the design, such as ultra-low power, small silicon footprint, reduced communication bandwidth and high efficiency to avoid information loss. The information extraction chain typically involves signal amplification and conditioning, spike detection, and spike sorting to determine the spatial and time firing pattern of each neuron. ududThis thesis first provides a background to the origin and basic electrophysiology of these biopotential signals followed by a thorough review of the relevant state-of-the circuits and systems for facilitating the neural interface.ududWithin this context, novel front-end circuits are presented for achieving resource-constrained biopotential amplification whilst additionally considering the signal dynamics and realistic requirements for effective classification. Specifically, it is shown how a band-limited biopotential amplifier can reduce power requirements without compromising detectability. Furthermore through the development of a novel automatic gain control for neural spike recording, the dynamic range of the signal in subsequent processing blocks can be maintained in multichannel systems. This is particularly effective if now considering systems that no longer requiring independent tuning of amplification gains for each individual channel. This also alleviates the common requirement to over-spec the resolution in data conversion therefore saving power, area and data capacity.ududDealing with basic spike detection and feature extraction, a novel circuit for maxima detection is presented for identifying and signalling the onset of spike peaks and troughs. This is then combined with a novel non-linear energy operator (NEO) preprocessor and applied to spike detection. This again contributes to the general theme of achieving a calibration-free multi-channel system that is signal-driven and adaptive. Another original contribution herein includes a spike rate encoder circuit suitable for applications that are not are not affected by providing multi-unit responses.ududFinally, spike sorting (feature extraction and clustering) is examined. A new method for feature extraction is proposed based on utilising the extrema of the first and second derivatives of the signal. It is shown that this provides an extremely resource-efficient metric than can achieve noise immunity than other methods of comparable complexity. Furthermore, a novel unsupervised clustering method is proposed which adaptively determines the number of clusters and assigns incoming spikes to appropriate cluster on-the-fly. In addition to high accuracy achieved by the combination of these methods for spike sorting, a major advantage is their low-computational complexity that renders them readily implementable in low-power hardware.