Classification of simple stimuli based on detected nerve activity

Describes an interface and a signal processing methodology that use measured neural signals to image overall axonal activity in intact peripheral nerve. Since one of the goals of this research is to create an interface that can eventually be used in both healthy and injured persons, an interfacing methodology that does not rely on nerve transection had to be developed. A cuff electrode containing multiple pairs of differential detectors was used to explore the feasibility of using measured neural signals to image overall axonal activity in intact peripheral nerve. The minimally invasive neural interfacing system (MINIS) consists of four parts: an in vivo multielectrode nerve cuff placed around an intact ensheathed whole nerve, wavelet based signal processing, information-theoretic data summarization, and a cascade correlation neural network. The system was validated using the visual system of Limulus polyphemus (common horseshoe crab). In our application the implantation of the cuff electrode requires surgery to expose the nerve but does not require removal of the sheath and surrounding connective tissue, hence the term "minimally invasive." The trained network for a given specimen was very specific to the specimen-interface-nerve configuration on which the data used to build the training/testing sets originated. When the network becomes overfitted it performs increasingly well at identifying the activity that corresponds to the data on which it was trained while becoming worse with novel data. Though it's doubtful a given source could ever be at the exact centers of all four pairs in a hand-mode cuff, being near the centers impacts the SNR and thus the accuracy for that pattern. Thus far the results are encouraging; however, more work is needed before this system could be used to reliably drive a prosthesis or interact with a virtual environment.