In electrophysiological signal processing for intramuscular electromyography data (nEMG), single motor unit activity is of great interest. The changes of action potential (MUAP) morphology, motor unit (MU) activation, and recruitment provide the most informative part to study the nature causality in neuromuscular disorders. In practice, for a single nEMG recording, more than one motor unit activities (in the surrounding area of a needle electrode) are usually collected. Such a fact makes the MUAP discrimination that separates single unit activities a crucial task. Most neurology laboratories worldwide still recruit specialists who spend hours to manually or semi-automatically sort MUAPs. From a machine learning perspective, this task is analogous to the clustering-based classification problem in which the number of classes and other class information are unfortunately missing. In this paper, we present a feature analysis strategy to help better utilize unsupervised (i.e., totally automated) methods for MUAP discrimination. To that end, we extract a large pool of features from each MUAP. Then we select the top ranked candidates using clusterability scores as selection criteria. We found spectrograms of wavelet decomposition as a top-ranking feature, highly correlated to the motor unit reference and was more separable than existing features. Using a correlation-based clustering technique, we demonstrate the sorting performance with this feature set. Compared with the reference produced by human experts, our method obtained a comparable result (e.g., equivalent number of classes was found, identical MUAP morphology in each pair of corresponding MU class, and similar histograms of MUs). Taking the manual labels as references, our method got a much higher sensitivity and accuracy than the compared unsupervised sorting method. We obtained a similar result in MUAP classification to the reference.