Within the last five years, major advances in electronics and micro-electrode technologies has led to the exciting new field of "Brain Machine Interfaces". Although the term might bring to mind images of the Frankenstein monster, a better example is the cochlear implant, a series of electrodes inserted into cochlear nucleus (where signals from the inner ear are initially processed). The electrodes are driven by input from a microphone, and the system allows many otherwise completely deaf individuals to hear well enough to talk on the telephone. The audio demonstration includes sound processed in from 1 to 16 narrow frequency bands that simulate the quality of the perception provided by the implant. This is an example of an afferent BMI, one that provides input to the brain from the outside world.

Audio Demonstration

For more information...

For more information...

My lab has begun to extend this work to the much more complex interface with a behaving monkey. The work is supported by a four year, $3.2 million Biomedical Research Partnership grant from the National Institute of Neurological Disorders and Stroke. Most BMIs developed to date have had two fundamental limitations. First, there has been relatively little attempt made to control devices with mechanical properties like the actual limb. A wide range of basic research, including that from my own lab, suggests that the primary motor cortex (M1) is organized to activate muscles and generate forces that gradually accelerate and move the limb to a new position. Existing BMIs use the signals recorded from the brain to designate position directly, for example, of the cursor on a screen, or of a robotic limb. Although it is more difficult to implement, we anticipate that a BMI designed to produce forces rather than positions may prove more natural for subjects to learn to control. We will develop a variety of virtual limb models that will vary in their complexity and realism. A variety of linear and nonlinear "decoders" must be developed that will take the neuronal signals and convert them into the signals necessary to control the virtual limb. These approaches will all be evaluated in terms of how quickly and accurately monkey subjects can learn to control the virtual limb under varied conditions.

Second, all existing BMIs rely on visual feedback to guide the movements. However, natural movement requires an interplay between the transmission of motor commands and the receipt of sensory feedback in the form of visual, tactile, and proprioceptive information.    Although vision is important for planning, "proprioception", the innate sense of limb position, is far more important in the guidance and control of movement. Under quite rare circumstances, human patients can lose this proprioceptive sense. The loss is devastating. These patients are typically wheel-chair bound, their arm movements are quite inaccurate, and even simple movement requires great concentration. Therefore, in addition to the interface that is necessary to extract control signals from the brain, we plan to develop methods to supply feedback signals directly to the brain, by electrically stimulating the primary somatosensory cortex (S1). S1 corresponds to M1, in the sense that it is the area of the brain that normally receives and processes tactile and proprioceptive signals from the moving limb. For this feedback interface, different kinds of "encoders" will be developed, that receive information about the state (position and movement) of the virtual arm, and convert it into trains of stimulus pulses that will be delivered to electrodes implanted in S1.

The project is a large collaboration among researchers at a number of different institutions. At Northwestern, the group includes Mussa-Ivalidi as well as Sara Solla. Solla is an expert on neural network models and information theory, and will assist in the development and testing of the decodes and encoders. Other collaborators include Nicho Hatsopoulos at the University of Chicago, an expert in multi-electrode recordings and uni-directional BMIs, Andy Barto at the University of Massachusetts at Amherst, an expert in reinforcement and machine learning, and Andy Fagg at the University of Oklahoma, specializing in robotics, who will develop the various virtual limb models and decoders of cortical activity. Ranulfo Romo from the University of Mexico, is a leader in the study of perception and decision making processes in the somatosensory system. He will be instrumental in the development of the sensory encoders.