Calibrating parameters in the MISCAN model using a genetic algorithm
Microsimulation disease models can be used to perform economic evaluations to inform policy makers on the long term costs and health outcomes of cancer screening and surveillance strategies. The Erasmus University Medical Center of Rotterdam, department of Public Health, developed the MISCAN model. Some parameters in the model should be estimated since they are not (directly) observable. These parameters are estimated through calibration, which is a method to obtain parameter estimates such that model outcomes fit the observed data. Currently, the Nelder-Mead simplex method is used, but that method does not perform well. Therefore, we proposed a new parameter search strategy called genetic algorithm. Using survival of the fittest and randomized changes, genetic algorithms try to find new solutions. Recombining two solutions (parent solutions) from a population and tweaking the parent solutions with certain probability, the aim is to create new and even better solutions (child solutions) in terms of fitness. We used the MISCAN model for esophageal cancer. Using different ways to initialize the algorithm and performing multiple calibration runs, we concluded that the proposed parameter search strategy seems to perform better in terms of running time and root mean squared prediction error for calibrating three parameters and therefore, it is a good alternative for the Nelder-Mead simplex method.