BayCANN

BayCANN performs Bayesian calibration of complex simulation models by training artificial neural network (ANN) metamodels as surrogates to estimate the full joint posterior distribution of calibrated parameters.

Key Features:

• ANN metamodeling: Trains an artificial neural network on samples of model inputs and outputs to serve as a surrogate for the original simulation model.
• Bayesian calibration of metamodel: Calibrates the trained ANN metamodel to obtain the posterior joint distribution of calibrated parameters.
• Calibration target generation: Uses confirmatory simulation analysis to generate targets such as adenoma prevalence and cancer incidence based on "true" parameter values from the literature.
• Comparison with IMIS: Evaluates performance against incremental mixture importance sampling (IMIS) for recovering true posterior parameter estimates.
• Computational efficiency: Demonstrated reduced runtime in the provided example (15 minutes for BayCANN versus 80 minutes for IMIS) for computationally expensive simulations such as microsimulations.

Scientific Applications:

• Colorectal cancer natural history modeling: Applied to calibrate a colorectal cancer natural history model using adenoma prevalence and cancer incidence targets.
• Health decision sciences model calibration: Used for Bayesian calibration of complex models in health decision sciences where full joint posterior parameter distributions are required.
• Microsimulation calibration: Suited for calibrating computationally expensive microsimulations by substituting ANN metamodels for the full model.
• Confirmatory simulation analysis: Supports studies that generate synthetic targets from literature-derived parameter values to assess parameter recovery.

Methodology:

Train an artificial neural network metamodel on a sample of model inputs and outputs, then calibrate the trained ANN metamodel to obtain the posterior joint distribution of calibrated parameters.