Effects of spike sorting error on information content in multi-neuron recordings

In brain-machine interface (BMI) applications, multi-channel recordings are spike sorted before the single-unit spike trains are used in computational analysis to decode the neural response. Nevertheless, questions about the necessity and effectiveness of spike sorting remain. To address one of those questions regarding how sorting error affects the ability to use these neural recordings for BMI's, Shannon information theory was applied to spike trains simulated with random sorting error. Mutual information rate was found to decrease exponentially with spike sorting error, regardless of type, i.e. whether false negative or false positive. Less than 10% error could be tolerated before the information content dropped to half its maximum value with no error. Implications for BMI applications are discussed.