| A number of neurological disorders may lead to varying degrees of paralysis in humans. A Brain-Computer Interface (BCI) based on an array of microelectrodes chronically implanted in the primary motor cortex can be used as an augmentative communication and control device for the paralyzed, essentially bypassing the damaged motor pathway by directly accessing and translating volitional control signals from populations of neurons in the patient's brain. In this research, the constraints imposed by motor cortical physiology on an implantable BCI are studied, and novel automatic signal processing strategies for extracting information from the noisy neurophysiological signals are developed. Three specific issues were investigated.; First, the ability of paralyzed individuals to internally modulate central activity patterns when attempting to move paralyzed limbs is established using functional Magnetic Resonance Imaging (fMRI), a noninvasive imaging method. Activity maps that qualitatively and quantitatively resemble the normal homuncular representation of limbs are demonstrated in the primary sensorimotor cortex as well as in other motor areas including subcortical structures like the cerebellum.; Second, an efficient algorithm that is able to automatically determine the number of units present in the noisy neural signals and classify them was developed. The use of the pattern recognition method of mixture decomposition is studied, and shortcomings in the previous use of gaussian-based mixtures are demonstrated. Multivariate t-distributions are introduced as a more suitable alternative, and two different agglomerative algorithms that use the t-distribution noise model are developed.; Third, a new method for effectively extracting control signals from a population of noisy primary motor cortical neurons is developed using a twofold framework inspired by applied estimation theory. First, data recorded in behaving monkeys is used to develop an effective statistical model of how the motor cortical neuron represents information about movement in its noisy time-dependent activity patterns. This encoding model and the estimation method of Sequential Monte Carlo filters are then used to optimally reconstruct the movement from the activity of populations of motor cortical neurons. |
