Leveraging Generative AI for Interpretable Clinical Decision Making Through Causal Graphs.

Area of Science:

• Artificial Intelligence in Medicine
• Causal Inference
• Health Informatics

Background:

• Clinical AI adoption is hindered by a lack of interpretability.
• Generative AI offers potential for medical knowledge integration.
• Structural Causal Models (SCMs) are crucial for reliable causal inference.

Purpose of the Study:

• To develop a computational framework using generative AI to create interpretable SCMs for clinical applications.
• To enhance clinical decision support, quality improvement, and population health management.
• To bridge the interpretability gap in clinical AI for evidence-based medicine.

Main Methods:

• A case study using the Midwest Healthcare Conference Causal Diagram Challenge dataset.
• Comparison of transformer-based large language models (LLMs) against human performance.
• Target trial emulation to estimate COVID-19 treatment effects on mortality using SCMs.
• Benchmarking against published randomized controlled trial results (RECOVERY trial).

Main Results:

• AI-designed SCMs achieved >90% bootstrap coverage for most COVID-19 patient severity strata.
• Both AI and human models showed equivalent clinical plausibility and similar statistical performance.
• SCM-based approaches demonstrated significantly higher coverage (76-98%) than traditional methods (1-37%).

Conclusions:

• Interpretable SCMs generated by AI can facilitate reliable causal inference in clinical settings.
• The framework enables meaningful human-AI collaboration while maintaining methodological rigor.
• SCMs are a promising solution for enhancing the adoption and trustworthiness of clinical AI.