Genetic algorithm-based personalized models of human cardiac action potential

We present a novel modification of genetic algorithm (GA) which determines
personalized parameters of cardiomyocyte electrophysiology model based on
set of experimental human action potential (AP) recorded at different
heart rates. In order to find the steady state solution, the optimized
algorithm performs simultaneous search in the parametric and slow
variables spaces. We demonstrate that several GA modifications are
required for effective convergence. Firstly, we used a mutation operator,
based on Cauchy amplitude distribution along with a random direction in
the parametric space. Secondly, relatively large number of elite organisms
(6-10 % of the population passed on to new generation) was
required for effective convergence. Test runs with synthetic AP as input
data indicate that algorithm error is low for high amplitude ionic
currents (1.6±1.6% for IKr, 3.2±3.5% for IK1, 3.9±3.5% for INa, 8.2±6.3%
for ICaL). Experimental signal-to-noise ratio above 28 dB was required for
high quality GA performance. GA was validated against optical mapping
recordings of human ventricular AP and mRNA expression profile of donor
hearts. In particular, GA output parameters were rescaled proportionally
to mRNA levels ratio between patients. We have demonstrated that
mRNA-based models predict the AP waveform dependence on heart rate with
high precision. The latter also provides a novel technique of model
personalization that makes it possible to map gene expression profile to
cardiac function.