Faster Simulation for Colon-cancer Screening: a Meta-modeling approach
Colorectal Cancer poses a heavy burden on the population, which can be relieved through screening programmes. The cost-effectiveness of different screening programmes can be predicted through simulation software, such as MISCAN-Colon. However, running the software is often too computationally expensive to evaluate all scenarios of interest, which may leave optimal screening programmes unexplored. This study compares different strategies to develop a metamodel for MISCAN-Colon, which can assist in the faster evaluation of screening programmes. Towards that purpose, it compares different strategies for sampling data and different model architectures. Based on a simulation study of MISCAN-Colon, five sampling strategies, which vary in how they sample risk and screening, and 168 Neural Network architectures are proposed. All model architectures are evaluated for all sampling strategies, which yields an optimal architecture per sampling strategy. For each sampling strategy a Bayesian Neural Network is fitted to function as a metamodel, which is then used to make predictions on a universal test set. The performance of each metamodel is then evaluated in terms of speed, accuracy and uncertainty estimation. The results indicate that randomly sampling screening and sampling risk from the population distribution yields the most accurate metamodel, although more work is needed to make it usable in practice.