| A neuroprosthetic interface seeks to bypass a damaged connection between brain and body for patients suffering neurodegenerative disease or traumatic injury. Ideally it should be able to distinguish when, what, and how a user wishes to move. This thesis is focused on decoding the neural activity accompanying reaching movements. Reaching movements are an ideal paradigm for a neuroprosthetic interface for three reasons. First, in the space of all possible movements, reach trajectories are limited to the subspace spanned by the dimensions of the target of the movement. Second, in addition to the peri-movement neural activity regime accompanying the executed trajectory of movement, in reaches, a plan activity regime pre-encodes the target. Third, reaches are clearly discrete, and can be distinguished from periods during which no movement is intended.; These three properties enabled the following specific contributions. The first contribution was the development of a hidden Markov model (HMM) approach for determining the regime of neural activity---unrelated to movement, movement preparation, or movement execution. Using the HMM transitions between regimes could be detected with an accuracy of a few 10s of milliseconds. This represents the first principled treatment of this problem, and contrasts with the few previous limited and ad-hoc approaches. The second contribution was to use the HMM to accurately estimate the desired target of movement. Additionally, coupling the HMM with a model of stereotyped movements enabled a 50% reduction in the error of decoded arm trajectories when compared with a traditional linear-filter approach. The final contribution was to develop a model of reaching movements which allowed for the integration of information about the target of movement without the constraint on target number imposed by the HMM implementation. Using a simple linear Gaussian model, we demonstrate significant reduction in the error of decoded movements without any target constraints. Furthermore, combining our two models, we demonstrate the previously unrealized ability to integrate information decoded from both preparatory and peri-movement neural activity. |
