Improving inversion of model parameters from action potential recordings with kernel methods

Abstract

AbstractCurrent methods for solving inverse problems in cardiac electrophysiology are limited by their accuracy, scalability, practicality, or a combination of these. In this proof-of-concept study we demonstrate the feasibility of using kernel methods to solve the inverse problem of estimating the parameters of ionic membrane currents from observations of corresponding action potential (AP) traces. In particular, we consider AP traces generated by a cardiac cell action potential model, which mimics those obtained experimentally in measurable in vitro cardiac systems. Using synthetic training data from the 1977 Beeler-Reuter AP model of mammalian ventricular cardiomyocytes, we demonstrate our recently proposed boosted kernel ridge regression (KRR) solver StreaMRAK, which is particularly robust and well-adapted for high-complexity functions. We show that this method is less memory demanding, estimates the model parameters with higher accuracy, and is less exposed to parameter sensitivity issues than existing methods, such as standard KRR solvers and loss-minimization schemes based on nearest-neighbor heuristics.