MedGuideX: Internalizing Decision Logic from Executable Guidelines into Large Language Models for Clinical Reasoning

Researchers introduce MedGuideX, a medical language model trained on executable clinical decision logic extracted from practice guidelines, achieving 10.28% accuracy improvement over existing methods. The approach transforms procedural guideline structures into synthetic training data that teaches models both correct clinical decisions and counterfactual reasoning, with physician validation confirming improved explanation quality.

MedGuideX represents a meaningful advancement in medical AI by addressing a fundamental limitation of current approaches: most systems treat clinical practice guidelines as unstructured text rather than exploiting their inherent procedural logic. This research demonstrates that the conditional decision trees embedded in CPGs—mapping patient variables to recommendations—contain learnable patterns that improve LLM clinical reasoning when properly formalized.

The breakthrough lies in the training methodology. Rather than fine-tuning on raw guideline text, the researchers reverse-engineered executable decision logic from guidelines, then generated synthetic question-answer pairs that capture both positive examples (correct guideline-aligned decisions) and counterfactuals (how recommendations change under different patient conditions). This approach mirrors how clinicians actually apply guidelines: by evaluating specific patient attributes against conditional criteria. The 10.28% relative accuracy gain across four benchmarks validates that this structured supervision translates to measurable performance improvements.

For the medical AI industry, this work addresses a critical trust deficit. Physician evaluation showing improved faithfulness, validity, completeness, and clarity of reasoning steps suggests MedGuideX produces not just more accurate outputs but more clinically defensible ones. This distinction matters significantly for clinical adoption, where model explainability and alignment with established medical standards directly impact liability and regulatory acceptance.

The framework is scalable and generalizable. Any clinical domain with documented practice guidelines—cardiology, oncology, infectious disease—can be transformed into training data using this pipeline. This creates a pathway for building specialized medical LLMs that genuinely internalize evidence-based decision logic rather than approximating it through statistical pattern matching.

• →MedGuideX converts procedural logic from clinical practice guidelines into synthetic training data that improves LLM accuracy by 10.28% on clinical reasoning tasks.
• →Physician evaluation confirms the model produces more faithful, valid, and complete clinical reasoning explanations compared to baseline medical LLMs.
• →The approach generates both factual and counterfactual training examples, teaching models how clinical decisions change under different patient conditions.
• →This methodology is generalizable across medical domains with established clinical practice guidelines, enabling scalable development of specialized medical AI systems.
• →Improved explainability and alignment with evidence-based standards addresses key adoption barriers for medical AI in clinical settings.