Learning Representations to Augment Statistical Analysis of Drug Effects on Nerve Tissues

We learn representations for classifying electrophysiological waveforms in a novel application, that is, analysis of electrical impulse conduction in microengineered peripheral nerve tissues. The goal is to understand the influence of distinct neuropathic conditions on the properties of electrophysiological waveforms produced by such tissues treated with known neurotoxic compounds and healthy controls. We show that statistical data analysis provides insight to the design of deep neural networks and the trained neural networks provide more informative representations compared to using pure statistical techniques. Based on this analysis, we design deep learning architectures that learn interpretable representations of the signals jointly with classification models. The proposed architecture provides new smooth representations of the signals that highlight the important points and patterns necessary for recognizing the electrophysiological effects of neurotoxic drug treatments.