BayCANN: Streamlining Bayesian Calibration With Artificial Neural Network Metamodeling

Purpose: Bayesian calibration is generally superior to standard direct-search algorithms in that it estimates the full joint posterior distribution of calibrated parameters. However, there are many barriers using health decision sciences stemming from need program complex models probabilistic programming languages and associated computational burden applying calibration. In this paper, we propose use artificial neural networks (ANN) as one practical solution these challenges. Methods: Calibration Artificial Neural Networks (BayCANN) involves (1) training an ANN metamodel on a sample model inputs outputs, (2) then calibrating trained instead language obtain We illustrate BayCANN colorectal cancer natural history model. conduct confirmatory simulation analysis by first obtaining parameter literature them generate adenoma prevalence incidence targets. compare performance recovering “true” values against performing directly incremental mixture importance sampling (IMIS) algorithm. Results: were able apply only dataset outputs minor modification BayCANN's code. example, was slightly more accurate true compared IMIS. Obtaining samples, running took 15 min IMIS which 80 min. applications involving computationally expensive simulations (e.g., microsimulations), may offer higher relative speed gains. Conclusions: uses distributions. Thus, can be adapted various levels complexity with or no change its structure. addition, efficiency especially useful models. To facilitate wider adoption, provide open-source implementation R Stan.