GROUND-TRUTH-BASED CHARACTERIZATION OF ACTIVITY-DEPENDENT SPIKE SHAPE VARIABILITY TO IMPROVE SPIKE SORTING PERFORMANCE

High-density microelectrode arrays (HD-MEAs) offer high spatiotemporal resolution and large sensing area. Each electrode typically records superimposed extracellular electrical potentials of multiple nearby neurons. Therefore, sophisticated downstream analysis pipelines (“spike-sorting”) are necessary to disentangle these superimposed signals to assess single-neuron activity.
Spike-sorting algorithms identify unique signal features that are characteristic of putative individual neurons (“units”). However, these features are assumed to be stationary over time and across different network states. This assumption has been challenged in recent studies, which show that the characteristic waveforms of individual neurons are highly variable, in particular during high-activity epochs, such as bursts.
Here, we combined patch-clamping with HD-MEAs to study waveform variations during network bursts and their implications on spike-sorting performance. We found strong changes in intracellular and extracellular waveforms during bursts. Based on ground-truth spike times, extracted from the patch-clamp data, we observed that spike-sorters miss a substantial fraction of spikes within bursts. We developed a mathematical model, that predicts activity-dependent extracellular waveform changes. Applying this model allowed us to rescue most of the spikes that were missed during initial sorting. Moreover, model parameters were variable across units and suggest a link to functional neuronal types. Our findings demonstrate that spike waveform characteristics depend on the recent firing history, and that such variations can lead to common spike-sorting failure modes. Studying the dynamics of waveform features enabled us to develop strategies to improve spike-sorting performance and could also be used to non-invasively infer underlying biophysical characteristics of individual neurons.