Information theoretic bounds on neural prosthesis effectiveness: The importance of spike sorting

We compute the capacity of neural prostheses using a vector Poisson process model for the neural population channel. For single-electrode stimulation prostheses, the capacity is proportional to the size of the population being stimulated, the same value that results when each neuron is stimulated individually. In contrast, when gross recordings are used in control prostheses, the capacity is much less than it is when each neuron's output is treated separately. Consequently, spike sorting, whereby gross recordings are sorted into their constituent spike trains, is crucial to the performance of neural control devices. By computing the capacity of the neural population channel with spike sorting, we find that false positives cause a far greater reduction in capacity than either missed spikes or mislabeled spikes. Thus, a good spike sorting algorithm for neural prostheses should be biased against committing false positives, even at the expense of altering the spike train statistics.