Leveraging Generative AI for Interpretable Clinical Decision Making Through Causal Graphs

Clinical AI systems' lack of interpretability limits their adoption in evidence-based medicine. To address this challenge, we propose a computational framework that harnesses generative AI's medical knowledge to create interpretable structural causal models (SCMs) for clinical decision support, quality improvement evaluation, and population health management. We evaluated our approach through a case study using data from the Midwest Healthcare Conference Causal Diagram Challenge, where we compared transformer-based large language models against human performance on a complex causal reasoning task: estimating COVID- 19 treatment effects through target trial emulation. Both groups designed SCMs to evaluate glucocorticoid treatment effects on 28-day mortality using real-world data from more than 2,000 hospitalized patients, benchmarked against published RECOVERY randomized controlled trial results. The best performing SCMs achieved bootstrap coverage rates exceeding 90% for two of three severity strata. Both human and AI models demonstrated equivalent clinical plausibility (n=3 expert reviewers) and similar statistical performance, though both struggled with critical disease severity. Ablation experiments comparing SCM-based approaches against traditional potential outcomes methods revealed SCMs achieved 76-98% coverage versus 1-37% for traditional methods. These results suggest that structural causal models can effectively bridge the interpretability gap in clinical AI by providing essential scaffolding for reliable causal inference and enabling meaningful human-AI collaboration while preserving methodological rigor essential for evidence-based medicine.