Adaptive Real-Time Decoding of Brain Signals for Long-Term Control of a Neuro-Prosthetic Device
Date of Award
Master of Science (MS)
Changes in the statistical properties of neural signals recorded at the brain machine interface (BMI) pose significant challenges for accurate long-term control of prostheses interfaced directly with the brain by continuously altering the relationship between neural responses and desired action. In this thesis, we develop and test an adaptive decoding algorithm that can recover from changes in the statistical properties of neural signals within minutes. The adaptive decoding algorithm uses a Kalman filter as part of a dual-filter design to continuously optimize the relationship between the observed neural responses and the desired action of the prosthesis. Performance of the algorithm was evaluated by simulating the encoding of arm movement by neurons in the primary motor cortex under stationary conditions as well as nonstationary conditions depicting loss and/or replacement of neurons in the population. The time taken for the system to fully recover (3-12 minutes) was faster than other adaptive systems (Rotermund et al 2006) and resulted in errors that were well matched to the initial system performance. The algorithm adapts to the local properties of the stimulus and is able to decode movements with high accuracy outside the trained movement space. This implementation lends itself favorably toward a portable, robust long-term decoding approach at the brain machine interface capable of providing accurate real-time decoding of neural signals over periods of weeks to months without outside intervention.