A Closed-Loop Framework for Knowledge-Based Causal Model Refinement Using Process Mining and Human-AI Governance

<title>Abstract</title> <p> <bold>Background:</bold> Directed acyclic graphs (DAGs) are a graph representation used in epidemiology for analysing causal inference, based on expert inputs and previous experiences. However, the lack of empirical evidence implies uncertainty in the causal structures, specially in chronic and complex disease with competing risks and temporal dynamics. This DAG uncertainity can bias the effect estimation regardless the methodological robustness. For this reason, we present a closed-loop framework that consider knowledge-based DAGs as scientific hypotheses, which are iteratively tested through governed human–AI collaboration loop based on process mining. <bold>Methods:</bold> Our framework comprises four different phases: i) generating a knowledge-driven event log using LLMs to produce a preliminary DAG from literature and expert input; ii) applying data-driven process discovery algorithms using the longitudinal event sequences as input data for producing an empirical direct follow graph (DAG); iii) iterating over the DAG refinement by comparing the empirical patterns in the DFG against the knowledge-based DAG, with human expert adjudication of each proposed change; and iv) statistical modeling and assessment using the final DAG. We operationalised the framework through an open-source interactive tool, and demonstrated its performance using synthetic data calibrated to a published cohort, comparing proton pump inhibitor (PPI) and histamine-2 receptor blocker (H2B) users. <bold>Results:</bold> After testing our framework against the synthetic dataset, the workflow reproduced longitudinal patterns in which CKD progression preceded cardiovascular adverse events, illustrating how the framework would surface a candidate mediator pathway for expert adjudication. Formal mediation analysis would be required to test the mediator hypothesis. Competing mortality rates illustrated how high mortality transitions would prompt expert review of competing-event handling in the analytical model. The AI tool recommended Fine--Gray competing-risk models, causal mediation analysis, and propensity-score balancing, all consistent with epidemiological best practice. <bold>Conclusions:</bold> The closed-loop framework provides a governed, transparent, and reproducible approach for integrating AI assistance into causal inference workflows. By embedding empirical appraisal into DAG construction while preserving human authority over causal interpretation, the framework advances methodological discipline in real-world evidence generation. </p>