Source-linked AI summary
Radiology Guideline-derived Decision Graph for Efficient Clinical Decision Support
Xuewen Zhou, Zhen Chen
TL;DR
Clinical guideline criteria are distributed across prose, tables, and definitions, while language models may deviate from guideline-based decisions. GDGraph compiles guidelines into executable graphs and combines localized System One judgments with deterministic traversal; across three radiology guidelines, it improves accuracy over the reported baselines at lower latency and cost.
Problem
Clinical guideline criteria are distributed across prose, tables, and cross-referenced definitions, and language models can deviate from established guidelines.
Source: §1 · p. 1Method
GDGraph compiles guidelines into decision graphs with linked definitions, then uses localized System One judgments and deterministic traversal to produce a traceable path.
Source: §1 · p. 1, §2.2 · p. 2Results
90.8% overall accuracy for GDGraph+Jev, versus 57.3% for Direct Jev, 80.1% for RAG+GPT-4o CoT, and 82.7% for GDGraph+GPT-4o CoT across three guidelines.
Source: §3.3 · p. 4Takeaways & Limitations
GDGraph’s explicit decision paths and linked definitions support auditable clinical decision support across the evaluated radiology guidelines.
Source: §4 · p. 4Errors included confusing “ill-defined margins” or “irregular shape” with “irregular margins” in 11 K-TIRADS cases and mapping qualitative vascularity descriptions to criteria in three O-RADS cases.
Source: §3.4 · p. 4
Abstract
from arXiv · showhide
Radiological risk stratification and management planning are essential downstream tasks in medical imaging. Automating these tasks is challenging because clinical guideline criteria are distributed across prose, tables, and cross-referenced definitions. We present Guideline-derived Decision Graph (GDGraph), a framework that decouples guideline compilation from case evaluation. During compilation, a large language model partitions guidelines into section-based chunks, extracts local decision subgraphs with linked definitions, and aggregates related branches into an executable decision graph. During inference, a fast System One model answers localized clinical questions while deterministic traversal resolves Boolean logic, turning opaque single-pass predictions into a traceable decision path without task-specific training. Across 347 cases spanning renal cyst, adnexal mass, and thyroid nodule guidelines, GDGraph improves System One accuracy from 57.3% to 90.8%, outperforming retrieval-augmented GPT-4o (80.1%) and graph-guided GPT-4o with chain-of-thought (82.7%) at substantially lower latency and cost. The source code will be available.
1. INTRODUCTION
Radiology guidelines make risk-stratification decisions through interdependent, discrete criteria, but language models can deviate from those guidelines. GDGraph organizes localized model judgments into a traceable decision process and improves accuracy across three guideline tasks.
Motivation
- Guideline-based risk stratification connects imaging findings to care through explicit, discrete criteria and successive decision steps.The tasks span renal cyst management, adnexal mass assessment, and thyroid nodule stratification.Source: §1 · p. 1
Problem
- Pretrained language models can process clinical reports by prompting without task-specific training, yet may deviate from established diagnostic guidelines.The central challenge is organizing local judgments while preserving conditional dependencies and producing a traceable outcome.Source: §1 · p. 1
Related Work
- RAG retrieves medical passages for model interpretation, while CPGPrompt converts segmented text into decision trees but relies on manual section selection and tree validation.Automatic construction remains challenging because criteria, thresholds, definitions, and exceptions are distributed across guideline documents.Source: §1 · p. 1
Contribution
- Across 347 radiology cases, GDGraph raises System One accuracy from 57.3% to 90.8%, above RAG (80.1%) and graph-guided GPT-4o with CoT (82.7%).The framework is presented as a practical approach to clinical decision support.Source: §1 · p. 1
2. METHODOLOGY
GDGraph compiles section-based guideline chunks into a unified decision graph with linked definitions, then uses localized System One judgments and deterministic traversal to reach a traceable decision.
2.1. Overview
- GDGraph separates offline guideline compilation from online case evaluation, using an LLM to chunk guidelines, retrieve linked definitions, extract local subgraphs, and aggregate related branches.The graph is constructed from guideline content that can include narrative text, figures, and tables.Source: §2.1 · p. 2
2.2. Guideline-derived Decision Graph
- Each graph contains clinical states, atomic conditions, directed transition rules, and candidate outcomes, with conditions linked to source-derived definitions.Rules combine atomic conditions using Boolean operators; section-based extraction keeps tables and footnotes intact, and exception clauses update affected rules while preserving other valid paths.Source: §2.2 · pp. 2–3
2.3. Graph-Guided Selection
- During online inference, a System One model answers active-node questions using attached definitions, and deterministic traversal follows satisfied branches to a final prediction.The traversal log records evaluated questions, answers, and edges as a traceable decision path.Source: §2.3 · p. 3
3. EXPERIMENTS
The experiments evaluate graph fidelity, decision performance, and inference efficiency across three radiology guidelines. GDGraph+Jev achieves the strongest overall decision performance, while error analysis identifies terminology and description-mapping challenges.
3.1. Experimental Setup
- The evaluation covers 347 text-based radiology cases across three clinical guidelines, with reference labels sourced or derived from guideline criteria.Case descriptions were reformulated to remove explicit diagnostic labels and reduce answer-leakage risk.Source: §3.1 · p. 3, Table 1 · p. 3
3.2. Graph Fidelity and Comparison with Retrieval
- 99.8% overall node precision and 96.4% rule precision indicate high source fidelity in the constructed graphs.Unstructured RAG+GPT-4o CoT raises overall GPT-4o CoT accuracy from 51.9% to 80.1%, while GDGraph+GPT-4o CoT reaches 82.7%.Source: §3.2 · p. 3, Table 2 · p. 3
- RAG performs better on O-RADS (84.6% vs. 78.6%), while graph guidance performs better on K-TIRADS (79.6% vs. 67.0%).K-TIRADS stratification jointly depends on nodule composition, echogenicity, and suspicious features.Source: §3.2 · p. 4, Table 3 · p. 4
3.3. System One Performance and Inference Efficiency
- 90.8% overall accuracy for GDGraph+Jev exceeds Direct Jev (57.3%), GDGraph+GPT-4o CoT (82.7%), and RAG+GPT-4o CoT (80.1%).It achieves the highest Acc., WF1, and MCC in each guideline; online API time is 0.79–1.62 s/case and cost is 0.56–1.07 × 10−3 USD/case.Source: §3.3 · p. 4, Table 3 · p. 4, Figure 2 · p. 4
3.4. Error Analysis
- In 11 K-TIRADS cases, related margin descriptions were conflated; three O-RADS errors involved mapping qualitative vascularity descriptions to node criteria.The analysis identifies normalizing radiology descriptions and aligning imaging concepts with model representations as important challenges.Source: §3.4 · p. 4
4. CONCLUSION
The conclusion presents GDGraph as a framework for converting radiology guidelines into executable decision graphs. Its experiments report higher accuracy than CoT-based GPT-4o baselines at lower latency and cost.
4. CONCLUSION
- GDGraph converts radiology guidelines into executable decision graphs for risk stratification and management planning.Compilation uses section-based chunking, local subgraph extraction, and cross-branch aggregation.Source: §4 · p. 4
- During inference, System One answers localized graph questions to produce a traceable decision path.Source: §4 · p. 4
- Across three guidelines, GDGraph+Jev achieves higher accuracy than CoT-based GPT-4o baselines at lower latency and cost.The conclusion describes this as indicating practical value for auditable clinical decision support.Source: §4 · p. 4
5. COMPLIANCE WITH ETHICAL STANDARDS
The study used human-subject data from the PubMed Central Open Access Subset and reports that ethical approval was not required under the attached license.
5. COMPLIANCE WITH ETHICAL STANDARDS
- The research was conducted retrospectively using human-subject data made available in the PubMed Central Open Access Subset.Source: §5 · p. 5
- The authors report that ethical approval was not required, as confirmed by the data license.Source: §5 · p. 5
- The data source is the PubMed Central Open Access Subset.Source: §5 · p. 5