Hardware-efficient data compression in wireless intracortical brain-machine interfaces

Abstract

Brain-Machine Interfaces (BMI) have emerged as a promising technology for restoring lost motor function in patients with neurological disorders and/or motor impairments, e.g. paraplegia, amputation, stroke, spinal cord injury, amyotrophic lateral sclerosis, etc. The past 2 decades have seen significant advances in BMI performance. This has largely been driven by the invention and uptake of intracortical microelectrode arrays that can isolate the activity of individual neurons. However, the current paradigm involves the use of percutaneous connections, i.e. wires. These wires carry the information from the intracortical array implanted in the brain to outside of the body, where the information is used for neural decoding. These wires carry significant long-term risks ranging from infection, to mechanical injury, to impaired mobility and quality of life for the individual. Therefore, there is a desire to make intracortical BMIs (iBMI) wireless, where the data is communicated out wirelessly, either with the use of electromagnetic or acoustic waves.

Unfortunately, this consumes a significant amount of power, which is dissipated from the implant in the form of heat. Heating tissue can cause irreparable damage, and so there are strict limits on heat flux from implants to cortical tissue. Given the ever-increasing number of channels per implant, the required communication power is now exceeding the acceptable cortical heat transfer limits. This cortical heating issue is hampering widespread clinical use. As such, effective data compression would bring Wireless iBMIs (WI-BMI) into alignment with heat transfer limits, enabling large channel counts and small implant sizes without risking tissue damage via heating.

This thesis addresses the aforementioned communication power problem from a signal processing and data compression perspective, and is composed of two parts. In the first part, we investigate hardware-efficient ways to compress the Multi-Unit Activity (MUA) signal, which is the most common signal in modern iBMIs. In the second and final part, we look at efficient ways to extract and compress the high-bandwidth Entire Spiking Activity signal, which, while underexplored as a signal, has been the subject of significant interest given its ability to outperform the MUA signal in neural decoding. Overall, this thesis introduces hardware-efficient methods of extracting high-performing neural features, and compressing them by an order of magnitude or more beyond the state-of-the-art in ultra-low power ways. This enables many more recording channels to be fit onto intracortical implants, while remaining within cortical heat transfer safety and channel capacity limits.