In numerous biological applications, including cardiac electrophysiology, the system of interest is studied by monitoring some quantities, called biomarkers in the following, that are extracted from the experimental measurements. These biomarkers are supposed to contain information about some hidden quantities which may be seen as uncertain parameters of a mathematical model. The question we will try to address in this talk is the following: is it possible to find "optimal" biomarkers for a given task (parameter estimation, classification, etc.)?
We propose a method to derive so-called composite biomarkers which is based on the resolution of a sparse optimization problem. We apply the method to several cardiac electrophysiology scenarios with both synthetic and experimental data. In particular, we address the problem of drug safety pharmacology using Micro-Electrode Array (MEA) measurements. We carry classification studies based on the drugs channel block effects and their arrythmogenicity using Machine Learning techniques. The classification scores improvements attributable to the use of composite biomarkers are discussed.