Optimizing Algorithms for Brain-Machine Interfaces

Imagine waking up trapped in a prison of your own flesh, blinking awake in the dull glow of a softly bleached hospital room. Your arms and legs are unresponsive to the will to move them, to the simple desire to reach up and scratch the itchiness of morphine from your eyes. Nothing happens, nothing responds, nothing moves, nothing feels. You are an immobile head trapped on an unresponsive body, and no matter how loudly you scream against the walls of your confinement from inside your head, nothing happens.

Luckily many quadriplegics retain the ability of speech and independent respiration. However, their quality of life, and that of the more unfortunate patients who are fully lucid but cannot communicate with the outside world by anything more than eye-blinking Morse code, remains severely compromised due to poor rehabilitation prospects and dependency upon caretakers. Perhaps one day we'll be able to repair injured nerves and restore connectivity with the rest of the body, but for now we're beginning to work out how to directly interface the brain with mechanical actuators. The goal is to develop an interface through which a paralyzed patient could control a machine that would augment their standard of living using nothing more than their mind and some technology.

Nerves and computers and both conveniently electrical. The problem is that they each operate on very different electrical schema. With computers we more or less know which transistor is storing which bit of information and have discrete units of conductors, resistors, capacitors and such. But it is very difficult to isolate the behavior of a single neuron in situ, so instead we use mass behavior of relatively large numbers of neurons to try to approximate functional firing chains. To translate the chaotic, non-discrete signal from neurons into discrete signals for machines, we need an algorithm. And to develop that algorithm, we need electrode-implanted animal models (unrestrained monkeys in this case) and scientists who can do math.

Li and Nicolesis et al have developed a novel modification of pre-existing brain-machine interface algorithms in "Unscented Kalman Filter for Brain-Machine Interfaces". Previous brain-machine interface algorithms were either linear approximations or non-linear particle filters. Linear filters (such as the Wiener and standard Kalman filters) didn't do as well approximating the behavior of neurons as particle filters, but they were much faster and computationally cheap enough that they could be used in real time. In general, particle filters (such as SSPPF) were very good at modeling the behavioral input of neurons, but they were so cumbersome that each iterative point required computing time that put it outside the range of useful real-time applications. Li and Nicolesis et al have modified a Kalman filter such that it uses a nonlinear tuning method (a quadratic model instead) and adds historical regression with multiple time offsets to help predict future behavior patterns. The quadratic tuning model integrates previous neural activity to arm movement models (the cosine tuning model, tuning to speed, tuning to distance of reach) into one cohesive and flexible equation. The historical regression with time offsets gives the model a short memory that allows it to quickly predict possible future states of neural activity to arm movement correlates and optimize them based on current neural activity input.

Implementing the quadratic neural tuning model greatly improved the accuracy of predicted neuron firing behavior, which means that the algorithm was better at interpreting the noisy signals from the neurons. At the same time, the historical regression made brain-activity-guided movement of the prosthetic hand* smoother and also helped to tune the monkeys' training on the algorithm by operating as a sort of continuous optimization that kicked out inefficient processes to improve overall performance. This quadratic model also significantly improved the reconstruction of the monkeys' desired hand movements during real-time tests in comparison to standard Kalman and Wiener filters.

In effect, this new algorithm construction allows for more accurate control of a prosthesis, with low tuning demands and progressive learning of efficient movement correlates. Ideally, this will one day allow paralyzed patients to accurately control mouse cursors (better than existing technology, anyway) or robotic prostheses that will greatly improve their quality of life. This technology may still be crude, but it is progressing rapidly and with great potential. I, for one, welcome the prospect of an auxiliary robotic arm. It would be great for my benchwork productivity.

*In this case, a cursor on a computer screen.

Li, Z., O'Doherty, J., Hanson, T., Lebedev, M., Henriquez, C., & Nicolelis, M. (2009). Unscented Kalman Filter for Brain-Machine Interfaces PLoS ONE, 4 (7) DOI: 10.1371/journal.pone.0006243

___________________________________________________


This is my entry in the Scientists' Duel that Hermitage and I are fighting for the title of Most Nefarious. Her entry is here. You, dear reader, will decide who wins. As of 12:00AM, 7/24/09, you have 72h to vote. You get 100 points to divide between Hermitage and I as you see fit. Report your scoring in comments. At the end of 72h we will tally up the points and determine the winner.