A computational model linking stochastic neural innervation processes and functional neuromuscular excitation is developed to investigate peripheral nerve interface based limb prostheses. A means of classifying the virtual nerve data is presented by using both a time domain feature set and a spike detection algorithm. Some intrinsic parameters in recording and classification, such as brachial fiber activation, analysis window length and feature selection, are discussed to achieve good neural signal recognition. Recommendations for optimal performance are made, with regard to information content and window length.