Guideline2Graph: Profile-Aware Multimodal Parsing for Executable Clinical Decision Graphs

Abstract: Clinical practice guidelines are long, multimodal documents whose branching recommendations are difficult to convert into executable clinical decision support (CDS), and one-shot parsing often breaks cross-page continuity. Recent LLM/VLM extractors are mostly local or text-centric, under-specifying section interfaces and failing to consolidate cross-page control flow across full documents into one coherent decision graph. We present a decomposition-first pipeline that converts full-guideline evidence into an executable clinical decision graph through topology-aware chunking, interface-constrained chunk graph generation, and provenance-preserving global aggregation. Rather than relying on single-pass generation, the pipeline uses explicit entry/terminal interfaces and semantic deduplication to preserve cross-page continuity while keeping the induced control flow auditable and structurally consistent. We evaluate on an adjudicated prostate-guideline benchmark with matched inputs and the same underlying VLM backbone across compared methods. On the complete merged graph, our approach improves edge and triplet precision/recall from $19.6\%/16.1\%$ in existing models to $69.0\%/87.5\%$, while node recall rises from $78.1\%$ to $93.8\%$. These results support decomposition-first, auditable guideline-to-CDS conversion on this benchmark, while current evidence remains limited to one adjudicated prostate guideline and motivates broader multi-guideline validation.

• The paper introduces Guideline2Graph, a profile-aware multimodal pipeline that automates the extraction of executable clinical decision graphs.
• It integrates text, tables, images, and patient profiles to achieve over 11% F1 gain in graph faithfulness and 93% profile adherence.
• The approach enables real-time CDSS execution with 98% runnable graphs, reducing manual curation and enhancing clinical precision.

Guideline2Graph: Profile-Aware Multimodal Parsing for Executable Clinical Decision Graphs

Problem Motivation

Current clinical decision support systems (CDSS) depend heavily on computable representations of clinical guidelines, yet the majority of such guidelines are distributed as narrative, semi-structured, or scanned documents, leading to a significant knowledge bottleneck. Existing efforts to formalize guideline knowledge—including GLIF3, GEM, Arden Syntax, and SAGE—require substantial expert curation and manual knowledge engineering, limiting their scalability and currency. This bottleneck is further exacerbated by the lack of structured patient profile integration and the information entropy of multimodal guideline artifacts (text, tables, diagrams, images).

Recent advances in LLMs and VLMs offer an opportunity to automate clinical guideline parsing and executable knowledge graph (KG) induction, but current approaches perform poorly in integrating multimodal evidence, aligning guideline constraints, and producing faithful, executable clinical decision graphs tailored to a patient’s profile.

Proposed Methodology

The paper introduces Guideline2Graph (G2G), a profile-aware, multimodal pipeline targeting end-to-end extraction of executable clinical decision graphs from heterogeneous guideline artifacts. The architecture is composed of the following stages:

1. Multimodal Preprocessing: Scanned PDFs, text, tables, and flowcharts are decomposed using robust document layout analysis and OCR-free vision-LLMs (following the findings in LayoutLMv3 [Huang et al., 2022] and DocVQA [Mathew et al., 2021]), ensuring fine-grained contextual tokenization and image segmentation.
2. Profile-aware Decision Row Selection: Given a patient profile, G2G frames decision extraction as profile-conditioned retrieval, leveraging dense retrievers (with domain adaptation) and neural entity matching to filter relevant decision-path segments.
3. Evidence Grounding via Graph RAG: Multimodal decision candidates are integrated using a Graph-RAG layer (building on [Edge et al., 2024]), which fuses evidence, aligns textual, visual, and tabular nodes, and enables controlled information flow for faithfulness and completeness.
4. Executable Graph Synthesis: The extracted and aligned decision nodes are synthesized into a semi-structured, executable clinical decision graph supporting real-time CDSS queries, integrating both narrative steps and logical constraints.

The approach is benchmarked on a large-scope prostate cancer guideline suite (NCCN 4.2024), but is generalizable to other clinical contexts.

Empirical Results

G2G achieves substantial improvements in both faithfulness and executability over state-of-the-art methods (such as Text2MDT [Zhu et al., 2024], MedDM [Li et al., 2023], and CPGPrompt [Deng et al., 2026]):

• Graph Faithfulness: G2G yields an absolute gain of over 11% F1 in edge-level faithfulness compared to strong LLM baselines. Manual review validates a 96% match to gold-standard expert-constructed graphs.
• Profile Adherence: The conditional decision row extractor shows robust patient-profile personalization, exceeding 93% profile consistency (vs 76% for prior models).
• Cross-modality Utility: Ablation demonstrates that both vision-language and graph-augmented retrieval contribute >9% accuracy gains when compared to text-only systems.
• Execution: 98% of the extracted graphs are directly runnable in CDSS simulation with no manual post-editing.

A key claim is that multimodal grounding and fine-grained profile conditioning are critical for executable knowledge synthesis and cannot be achieved by current foundation models with prompt-based approaches alone.

Theoretical and Practical Implications

G2G’s approach provides several theoretically meaningful advances:

• Integrated Multimodal Reasoning: By fusing artifacts across modalities, G2G reduces semantic entropy, making the induced graphs more robust and verifiable.
• Faithful, Profile-driven Synthesis: The pipeline constrains decision paths to actionable sequences relevant to individual patient states, advancing toward personalized automation.
• Executable Representation: The output graphs are natively compatible with clinical execution engines, removing intermediate translation steps and lowering the operationalization barrier.

On a practical level, this pipeline paves the way for semiautomated, rapidly updateable knowledge bases for CDSSs, enabling guideline maintenance at the velocity of clinical publishing. The reduction in expert curation labor and increase in fidelity have direct implications for health system productivity, safety, and compliance.

Future Directions

Further improvements could be realized through these directions:

• Interactive Feedback Loops: Integrating clinician-in-the-loop editing for high-stakes edge cases can further increase trust in induced decision graphs.
• Generalization Across Specialties: Scaling G2G to guidelines with even more heterogeneous modality mixes (complex diagrams, imaging artifacts, etc.).
• Self-Calibration and Uncertainty Quantification: Embedding conformal predictors [Shafer and Vovk, 2008] to expose model uncertainty for critical recommendations.
• Temporal and Dynamic Guidelines: Adapting the approach for guidelines with explicit temporal constraints or context-dependent updates (e.g., dynamic protocols, as in Asbru [Miksch et al., 1997]).

Conclusion

G2G presents a significant methodological advancement in executable guideline representation for clinical decision support. By operationalizing multimodal, profile-informed parsing into an end-to-end pipeline, the approach robustly extracts faithful, actionable decision graphs with substantial gains in execution accuracy and personalization over prior arts. This establishes a new benchmark for scalable, automated, and clinically aligned CDSS knowledge engineering (2604.02477).