Review of machine learning and signal processing techniques for automated electrode selection in high-density microelectrode arrays

Abstract

Recently developed CMOS-based microprobes contain hundreds of electrodes on a single shaft with interelectrode distances as small as 30 µm. So far, neuroscientists manually select a subset of those electrodes depending on their appraisal of the "usefulness" of the recorded signals, which makes the process subjective but more importantly too time consuming to be useable in practice. The everincreasing number of recording electrodes on microelectrode probes calls for an automated selection of electrodes containing "good quality signals" or "signals of interest." This article reviews the different criteria for electrode selection as well as the basic signal processing steps to prepare the data to compute those criteria. We discuss three of them. The first two select the electrodes based on "signal quality." The first criterion computes the penalized signal-to-noise ratio (SNR); the second criterion models the neuroscientist's appraisal of signal quality. Last, our most recent work allows the selection of electrodes that capture particular anatomical cell types. The discussed algorithms perform what is called in the literature "electronic depth control" in contrast to the mechanical repositioning of the electrode shafts in search of "good quality signals" or "signals of interest."

Neural probes – microsystems to interface with the brain

“…On the other hand electrodes were actively selected by complementary metal-oxide-semiconductor (CMOS)-based electronics integrated in the slender probe shafts [5]. Specific machine learning algorithms were further applied to identify those electrodes out of hundreds of recording sites along each single probe shaft [20] that are most appropriate for dedicated experiments in different brain regions [5].…”

Defining the biological and clinical basis of radiomics

“…Machine learning comprises a large panel of statistical algorithms that aim at fitting coefficients of classification models using training data in order to provide accurate predictions on novel data [68]. Machine learning is widely employed in biomedical sciences to develop predictive and prognostic biomarkers, especially in the field of gene expression signatures [69] as a majority of machine learning methods require quantitative data.…”

“…2) Segmentation: Delineation or segmentation of the three dimensional tumor volume is the basis of extracting radiomic tumor imaging features, although historically features were also extracted from two dimensional slides [66][67][68]. Segmentation of tumors is usually performed on axial slices following the image acquisition of radiographic modalities.…”

“…From all features, the most informative are selected with unsupervised and supervised feature selection procedures [63]. A number of current studies also employ a stability dataset for preprocessing to filter for highly reproducible features [49,68,76]. As all features are quantitative, statistical models can be fitted to test for cor-19 relations and predictive power for clinical endpoints, including overall and progressionfree survival [64].…”