Application of k-Nearest Neighbors Method for Drug Concentration and Cardiotoxicity Classification Using Extracellular Field Potentials and Reconstructed Action Potentials of Cardiac Cells

Abstract

Micro-electrode array (MEA) systems are important for measuring extracellular field potentials (FP) of cardiac cells, which is a crucial step in cardiotoxicity assessment. However, without modification, the MEA system is only capable of recording FPs. This limits the number of parameters for cardiotoxicity assessment only to FP parameters, while the action potential (AP) parameters remain unused. To address this issue the MEA systems are often modified to use electro- or optoporation to record the local extracellular APs (LEAPs), which allows to reliably quantify the AP morphology. As an alternative to MEA modification and cell membrane stimulation the AP can be reconstructed mathematically.This study explores how using additional parameters from reconstructed action potentials (RAPs), derived from FPs, can improve the accuracy of k-NN machine learning models for drug concentration and potential cardiotoxicity classification. The k-NN classifier was trained using combinations of FP and RAP parameters. The k-NN models were evaluated using five-fold stratified cross-validation and cross-channel validation. Their performances were compared using error rate, macro precision, macro recall and macro F1 score accuracy metrics. The results indicated that ncorporating RAP parameters into the feature set increased the F1 score of k-NN model for DMSO concentration classification by up to 10.78% compared to the training set with only FP features.