Automated Neurons Recognition and Sorting for Diamond Based Microelectrode Arrays Recording: A Feasibility Study

Microelectrode arrays (MEA) are extensively used for recording and stimulating neural activity in vitro and in vivo. Depositing nanostructured boron doped diamond (BDD) onto the neuroelectrodes makes it possible to obtain dual mode low-noise neuroelectrical and neurochemical information simultaneously. The signal processing procedure requires finding and distinguishing individual neurons spikes in the recordings. Spike identification is usually done manually which is inaccurate and inappropriate for complex datasets. In this paper, we present a methodology and two algorithms for neurons recognition and evaluation based on unsupervised learning. Forty-five extracellular randomly selected signals from 26 unique measurements of embryonic hippocampal rat neurons (20 kHz, 6 min) were recorded on the commercial 60 TiN channels MEA. The signals were filtered in the 300–3000 Hz band and an amplitude detector (4x std of the background noise) was used for spike detection. WaveClus features were computed and its 3 PCA components were extracted for every spike. The optimal number of clusters were evaluated by an expert rater. K-means + gap criterion (alg. 1) and the Gaussian Mixture Model + Bayesian Information Criterion (alg. 2) were implemented and compared. The total IntraClass Correlation showed a significant inter-rater agreement for all 3 rater procedures (ICC = 0.69, p < 0.001), when post hoc weighted Cohen’s Kappas for 2 raters were 0.85 (expert vs. alg. 1; p < 0.001) and 0.62 (expert vs. alg. 2; p < 0.001). This will contribute to the objective definition of dual mode BDD MEA performance criteria and for a comparison with the current system.