Title: Deep Learning Architectures for FSCV, a Comparison

Abstract: We examined multiple deep neural network (DNN) architectures for suitability in predicting neurotransmitter concentrations from labeled in vitro fast scan cyclic voltammetry (FSCV) data collected on carbon fiber electrodes. Suitability is determined by the predictive performance in the "out-of-probe" case, the response to artificially induced electrical noise, and the ability to predict when the model will be errant for a given probe. This work extends prior comparisons of time series classification models by focusing on this specific task. It extends previous applications of machine learning to FSCV task by using a much larger data set and by incorporating recent advancements in deep neural networks. The InceptionTime architecture, a deep convolutional neural network, has the best absolute predictive performance of the models tested but was more susceptible to noise. A naive multilayer perceptron architecture had the second lowest prediction error and was less affected by the artificial noise, suggesting that convolutions may not be as important for this task as one might suspect.