Home // International Journal On Advances in Life Sciences, volume 13, numbers 1 and 2, 2021 // View article
Authors:
Benjamin Schindler
Dorothee Günzel
Thomas Schmid
Keywords: Physiology; Epithelia; Impedance Spectroscopy; Machine Learning; Least Squares; Neural Networks; Random Forests.
Abstract:
Quantifying tissue permeability is a central task in assessing pathophysiology of intestinal epithelia. A common and convenient approach for this task is to determine electric properties like resistance and capacitance of the epithelial tissue by applying impedance spectroscopy. While the measurement technique itself is well-established, analysis tools and strategies are still subject to ongoing research in epithelial physiology. Estimations of electric parameters are known to be particularly imprecise for models where apical and basolateral sides of the tissue differ significantly from each other. One-sided application of substances such as Nystatin play an important role here, as they alter membrane conductivity on one side of the tissue while leaving other properties unchanged. Here, we present a novel method that considers two functional states of the cells, namely before and after apical addition of the substance Nystatin. To this end, an extensive dataset modeled after the epithelial cell lines HT-29/B6, IPEC-J2, and MDCK I was synthesized. In a broad study, we show that considering features from two distinct tissue states leads to significantly better regressions by decision trees, random forests, and multilayer perceptrons. Therein, we extend previous work in order to progress from a two-path to a more revealing three-path model of electric tissue properties. Parameters of a corresponding equivalent circuit could be determined with less than five percent deviation from the known target value on average. In a post-processing step, predictions by independent machine learning regressions are employed to initialize a least squares parameter fitting, here the associated impedance spectrum is aligned with the originally observed spectrum, reducing the residual sum of squares by 99% on average.
Pages: 134 to 148
Copyright: Copyright (c) to authors, 2021. Used with permission.
Publication date: December 31, 2021
Published in: journal
ISSN: 1942-2660