Minimizing data transfer with sustained performance in wireless brain–machine interfaces

Abstract

Brain–machine interfaces (BMIs) may be used to investigate neural mechanisms or to treat the symptoms of neurological disease and are hence powerful tools in research and clinical practice. Wireless BMIs add flexibility to both types of applications by reducing movement restrictions and risks associated with transcutaneous leads. However, since wireless implementations are typically limited in terms of transmission capacity and energy resources, the major challenge faced by their designers is to combine high performance with adaptations to limited resources. Here, we have identified three key steps in dealing with this challenge: (1) the purpose of the BMI should be clearly specified with regard to the type of information to be processed; (2) the amount of raw input data needed to fulfill the purpose should be determined, in order to avoid over- or under-dimensioning of the design; and (3) processing tasks should be allocated among the system parts such that all of them are utilized optimally with respect to computational power, wireless link capacity and raw input data requirements. We have focused on step (2) under the assumption that the purpose of the BMI (step 1) is to assess single- or multi-unit neuronal activity in the central nervous system with single-channel extracellular recordings. The reliability of this assessment depends on performance in detection and sorting of spikes. We have therefore performed absolute threshold spike detection and spike sorting with the principal component analysis and fuzzy c-means on a set of synthetic extracellular recordings, while varying the sampling rate and resolution, noise level and number of target units, and used the known ground truth to quantitatively estimate the performance. From the calculated performance curves, we have identified the sampling rate and resolution breakpoints, beyond which performance is not expected to increase by more than 1–5%. We have then estimated the performance of alternative algorithms for spike detection and spike sorting in order to examine the generalizability of our results to other algorithms. Our findings indicate that the minimization of recording noise is the primary factor to consider in the design process. In most cases, there are breakpoints for sampling rates and resolution that provide guidelines for BMI designers in terms of minimum amount raw input data that guarantees sustained performance. Such guidelines are essential during system dimensioning. Based on these findings we conclude by presenting a quantitative task-allocation scheme that can be followed to achieve optimal utilization of available resources.