Machine Learning-Assisted Detection of Action Potentials in Extracellular Multi-Unit Recordings

Recording of neural activity with extracellular microelectrode arrays (MEAs) is a powerful technique for studying neuronal signalling in vivo. The activity of neurons, in the form of action potentials or spikes, is isolated from extracellular signals by means of two fundamental procedures: spike detection and spike sorting. Here we focus on the first problem, namely to identify the portions of signal that correspond to the firing of neurons. The most commonly used technique to identify spikes in extracellular signals is to set a hard threshold, either manually or automatically, based on the features of the signal, typically as a factor of the signal's standard deviation. This approach is simple, fast, and amenable for on-line spike detection, but arguably suboptimal because it ignores spikes with amplitudes just below the threshold. Here we developed a machine learning approach that is complementary to the threshold-based method with the aim of enriching spike detection in off-line analysis. We show that this approach has high accuracy with respect to threshold-based detection. More importantly, our results indicate that, using this approach, a significant amount of additional neuronal spikes can be identified that are ignored by the threshold method. We argue that classical spike detection, based on thresholding, misses out a significant fraction of real spikes, thereby strongly impacting the validity of spike-train analyses, and propose a method that can be used to alleviate this problem.