Background

A Brain Machine Interface is a connection between a brain and a device which enables neural signals (spikes) to direct an external activity. Its purpose is to restore movement and control for people with damaged sensory and motor functions. In order to arrive at a practical implementation of a BMI, the way the brain encodes and manipulates information should be thoroughly understood. A vital component of brain machine interface (BMI) instrumentation is the ability to detect the presence and timing of action potentials (spikes) in real-time from the electrode signal. BMIs typically use changes in the instantaneous firing rates of individual neurons to indicate how to control a prosthesis. Spike detection is of even more relevance for wireless BMI systems because per-channel data rates are dramatically reduced by only transmitting information when spikes are detected. Spike detection in wireless BMI systems must balance the need to detect as many spikes as possible (to minimize information loss) versus the need to minimize the number of false detections (which increase the per-channel data rate).

Spike Detection

The problem of detecting spikes in extracellular cortical electrode signals has been studied extensively. Having received far less attention is the problem of how best to detect spikes in embedded systems where clock cycles and battery power are at a premium, and detection must work in real-time. We have endeavored to address this issue in a number of ways. We have conducted a series of analyses to compare the performances of various spike detectors, with a focus on algorithms that can be implemented in real-time embedded systems. Our research has shown that simple detectors, such as taking the absolute value of the signal and applying a threshold, are often as effective as their more sophisticated counterparts. Our research has also shown that such algorithms can be implemented in commercially available programmable logic devices to do real-time detections in 96 channels simultaneously. In pursuing this work, we have also developed novel mathematical criteria for evaluating spike detector performance.

The Gold Standard Problem

An important and widely recognized problem in this area is known as the Gold Standard Problem. In order to evaluate the performance of a spike detector, one must know the true spike times. Unfortunately, when dealing with signals recorded from extracellular electrodes, true spike times are typically only estimates that are generated using spike detection software. Therefore, one spike detection estimate is compared against another, and not against “true” spike times. An alternative strategy is to use simulated data in which true spike times are known precisely. However this option is limited by the fact that it is extremely difficult to generate truly realistic synthetic signals.

One promising option investigated in our lab is the use of a special data set created by Professor Buszaki (Rutgers University) in which simultaneous extracellular and intracellular recordings were made. Extracellular spike detection can be accurately quantified since spike times can be precisely deduced from the intracellular traces. Using this data set, we have verified our early findings that simple thresholding and energy based detectors have comparable performance to more sophisticated detectors.

Research in this field would be greatly enhanced by the presence of a public library of extracellular neural signals with the associated spike times, as estimated by the investigator.

Performance Requirements

A vital question that has yet to be adequately addressed in the literature is what are the minimum requirements for a spike detector that operates in the context of adaptive cortical tissue. In a BMI application, the brain has been shown to continuously modify its behavior to optimize its ability to control the prosthesis. The question that should be posed by instrumentation engineers is not “what is the best spike detector” but rather “what is the minimum spike detector performance that the brain can adapt to and still adequately control the prosthesis.” In other words, the adaptive brain is a fault-tolerant system – how much loss of information can the brain tolerate before it loses its ability to adequately control the prosthesis. This logic should be applied to all other subsystems of BMI instrumentation: spike sorting, the number of channels, the number of bits of resolution, bit error rate, and buffer overrun. It is impossible to address these questions in an open-loop setting, that is, without taking into account the plasticity of the brain.

Time and ethical considerations make it unreasonable to test myriad instrumentation combinations with live subjects. Our lab is therefore developing a hybrid software – hardware model as a means for addressing this problem. The model will use an adaptive filter implementation of the Marr-Albus cerebellum model to control the trajectory of a robotic arm. A more complete description of this model can be read here.