Abstract for: Facing the Challenge of Modeling Personalized Medical Interventions: A Data-driven Workflow

System dynamics models (SDMs) are promising tools for personalized medicine when based on comprehensive causal loop diagrams (CLDs) calibrated with longitudinal data. We previously introduced an expert-driven approach to reduce uncertainties when converting such CLDs into SDMs. However, a more data-driven workflow for model selection can complement expert-solicitation where domain knowledge is scarce or conflicting. Here, we propose such a workflow, which includes causal structure discovery to mitigate uncertainty regarding CLD edges and, crucially, a model selection procedure to identify plausible (nonlinear) functional forms for these edges. The workflow also aims to improve predictive accuracy and model validity, two major challenges for the application of SDMs to (personalized) medicine. We conclude by proposing that this methodology becomes part of an inductive-deductive development cycle, where a collaboration between modelers and experimental researchers involves iteratively identifying and then reducing important sources of uncertainty. This workflow holds the promise to further refine SDMs beyond what can be achieved by expert solicitation or algorithmic causal discovery alone.x5xx