Can we hear the shape of a neuron? Cell type classification in high density multi-electrode recordings

Abstract

Event Abstract Back to Event Can we hear the shape of a neuron? Cell type classification in high density multi-electrode recordings Charl A. Linssen1* and Paul H. Tiesinga1 1 Donders Institute, Netherlands There is a trend toward the use of in vivo electrode probes with hundreds or even thousands of contact sites. The contact pitch (<50 μm) is much smaller than the spatial extent of a single neuron (~1 mm); thus, the extracellular potential (EP) generated by a single cell can be detected on many sites simultaneously [1]. Morphology of the cell determines the spatial profile of the EP, because most compartments of the cell support passive or active transmembrane currents [2]. We hypothesize that morphological characteristics of the neuron (cell type) can be inferred by inspecting this profile, where existing classification approaches rely exclusively on temporal information from a single electrode contact. We set up a software pipeline consisting of a forward model and classifier. Neurons with randomized morphologies were generated [3] according to two templates: pyramidal and multipolar (Fig. 1A). The model neurons had a passive membrane together with active Hodgkin-Huxley channels in the soma only. Cells received balanced excitatory and inhibitory synaptic inputs that were uniformly distributed along the dendritic tree. Extracellular potentials were calculated using Poisson's equation for currents under the assumption of an unbounded, homogeneously resistive extracellular medium [2][4]. The multiunit activity (MUA) was obtained by acausally band-pass filtering the EP between 0.3 and 3 kHz, after which the spatial profile at the peak value of the MUA was extracted (Fig. 1B). The activity of all other active cells in the area surrounding the electrode was modeled as additive white Gaussian noise, of which the standard deviation was varied. Sample profiles were concatenated and fed into an SVM classifier after pre-processing (Fig. 1C). Reported values (Fig. 1D) were obtained using 50/50 cross-validation and for SVM hyperparameter values optimized to minimize test error. In the absence of noise, morphological classes can be discriminated with a reliability of more than 90% for neurons at distances from the electrode up to 250 μm. Increasing noise amplitude pushed the classifier performance toward chance level, both decreasing in accuracy at close distances and restricting successful classification to a smaller range of distances between cell and electrode. A notable increase in classification error occurred at the boundary of the dendritic tree, corresponding to the boundary between the near and far electric field. We conclude that within the context of this model, classification of gross neuronal morphology based on the EP spatial profile is feasible. Proximity of the recording sites to the dendritic tree is not strictly necessary, as the far-field component could also be used successfully for classification. Real-world benchmarking of the classifier could benefit from experimental data with a known ground truth, for instance obtained using cell type-specific optogenetic activation. Figure 1 Acknowledgements The research leading to these results receives funding from the European Union's Seventh Framework Programme (FP7/2007-2013) under grant agreement no. 600925.