Personalization is the process of fitting a model to patient data, a criticalstep towards application of multi-physics computational models in clinicalpractice. Designing robust personalization algorithms is often a tedious,time-consuming, model- and data-specific process. We propose to use artificialintelligence concepts to learn this task, inspired by how human expertsmanually perform it. The problem is reformulated in terms of reinforcementlearning. In an off-line phase, Vito, our self-taught artificial agent, learnsa representative decision process model through exploration of thecomputational model: it learns how the model behaves under change ofparameters. The agent then automatically learns an optimal strategy for on-linepersonalization. The algorithm is model-independent; applying it to a new modelrequires only adjusting few hyper-parameters of the agent and defining theobservations to match. The full knowledge of the model itself is not required.Vito was tested in a synthetic scenario, showing that it could learn how tooptimize cost functions generically. Then Vito was applied to the inverseproblem of cardiac electrophysiology and the personalization of a whole-bodycirculation model. The obtained results suggested that Vito could achieveequivalent, if not better goodness of fit than standard methods, while beingmore robust (up to 11% higher success rates) and with faster (up to seventimes) convergence rate. Our artificial intelligence approach could thus makepersonalization algorithms generalizable and self-adaptable to any patient andany model.
展开▼
