Authors :
Presenting Author: Andrew Michalak, MD, MS – New York University
Edward Merricks, PhD – Associate Research Scientists, Epilepsy, Columbia University Irving Medical Center; Robert Goodman, Neurosurgeon – Neurosurgeon, Neurosurgery, New Jersey Brain and Spine; Sameer Sheth, MD, PhD – Neurosurgeon, Neurosurgery, Baylor College of Medicine; Neil Feldstein, MD – Neurosurgeon, Neurosurgery, Columbia University Irving Medical Center; Guy McKhann, MD – Neurosurgeon, Neurosurgery, Columbia University Irving Medical Center; Catherine Schevon, MD, PhD – Epileptologist, Epilepsy, Columbia University Irving Medical Center
Rationale:
Quantifying excitatory and inhibitory firing in microelectrode recordings is a commonly employed technique in epilepsy and cognitive research. However, current methods require sorting action potentials (APs) into single neurons, which is time-intensive and might discard potentially useful APs that were not assigned to identified single units. Since epilepsy research focuses on cell-type specific population activity, limiting APs to single units may introduce bias by focusing on a small number of high firing rate units. We present an unsupervised method to generate probabilistic cell-type specific firing rates based on unsorted AP detections that is significantly faster than manual sorting.Methods:
Data came from 33 epilepsy patients in which microelectrodes were included alongside invasive clinical electrodes (8 Utah array, 25 Behnke-Fried). 66 interictal segments were used (21 Utah array, 45 Behnke-Fried). All samples were analyzed by traditional spike sorting. Artifact was removed from raw detections using a novel toolbox built from line-noise harmonics and non-physiological waveforms. Next, leveraging the observation that AP half-width (HW) durations of excitatory and inhibitory cells should form discernable log-normal distributions, a two component Gaussian mixture model (GMM) was fitted to the log of the HWs. Probabilistic excitatory and inhibitory firing rates were generated by weighting the contribution of each AP by its posterior probability from the two components in the GMM. The performance of the GMM was evaluated by classifying APs from manually identified spikes from all recording segments. The probabilistic firing rates during a subset of four non-spike-sorted interictal epochs and one ictal epoch were compared to the single unit firing rate (FR) derived from the manually sorted data using Pearson correlation.
Results:
The method achieved high confidence for identifying excitatory waveforms from the manually sorted dataset (median confidence 100%, mean 95.7%, SD 17.5%) and modest confidence for predicting inhibitory waveforms (median confidence 81%, mean 61%, SD 41%). When correlating the probabilistic firing rate to the manually sorted FR, the excitatory firing rate had a mean correlation of 0.988 (range 0.983-0.991) and the inhibitory firing had a mean correlation of 0.907 (range 0.836-0.941). To explore if the method captures cell-type specific firing patterns during seizures, we tested it on an ictal recording with known out-of-phase inhibitory firing to the dominant ictal rhythm. The method captured both cell-type's firing patterns, showing no difference to the single unit approach and similarly detecting phase differences between the two (p < 0.001; circular k-test).