A Causal Discovery Workflow for Rare Diseases: Experts-in-the-Loop Analysis of Sparse Longitudinal Data

Causal networks provide a mechanistic understanding of clinical phenomena, allowing for personalized and explainable decision-making. Causal discovery, namely the task of constructing such models, is challenging, particularly for rare diseases, where observational data are sparse, medical knowledge is incomplete, and diseases develop over time. This work proposes a new and original expert-in-the-loop causal discovery workflow that iteratively refines a set of causal networks associated with different disease mechanisms. When applied to soft tissue sarcoma, a heterogeneous group of rare cancers, the workflow allows for the first comprehensive causal description of the disease's natural history. Indeed, three causal networks associated with different disease mechanisms shed light on the complex interplay between patients' covariates and disease behavior. These results have the potential to enhance clinical decision-making by allowing the development of personalized treatment strategies. The proposed workflow paves the way to agile, modular, and flexible causal discovery for clinical domains characterized by data sparsity, longitudinal dynamics, and heterogeneous expert knowledge.