Combined biophysical simulations and machine learning approaches are increasingly used to enhance and validate data analysis techniques in neuroscience. These approaches are favored in applications in which interpretations of the results of data analyses are limited by lack of ground-truth knowledge of the fundamental constraints of the biophysical system. Attempts at classifying neurons using recordings of extracellular action potentials (EAPs) have largely remained limited to binary classifications of putative excitatory and inhibitory types. We demonstrate a data-driven analysis of EAPs that addresses this limitation by demixing the contributions of various morphological sources to observed EAPs. This analysis aids in characterizing gross morphological properties of neuron-types and is validated using biophysically-realistic computational models of multiple neuron-types from rodent sensory cortices. Our approach facilitates automated feature extraction and selection for mapping EAPs to spatial patterns along morphological domains of different neuron-types. Multiple sets of features are used for classification based on the demixed EAP spatial patterns. These differentiating features span multiple timescales and reflect the degree of contribution of model compartment domains to the multiresolution features of simulated EAPs. Finally, we show how this data-driven approach can be applied to multichannel recordings of cortical neurons to enhance neuron-type classifications.