Clinical Guidance Trees (CGT)

• Clinical Guidance Trees are hierarchical, expert-derived decision protocols represented as rooted, directed trees that systematically deliver evidence-based clinical recommendations.
• Modern methodologies use end-to-end LLM generation and pipeline extraction to map clinical text into CGT structures, enhancing node, edge, and tree accuracy.
• CGTs integrate with clinical decision support systems through automated traversal, LLM-driven patient dialogue, and real-time error monitoring for improved care.

A Clinical Guidance Tree (CGT) formalizes hierarchical, expert-derived decision processes from clinical practice guidelines as rooted, directed tree structures with decision nodes and terminal recommendations. CGTs codify evidence-based protocols, segmenting patient populations by decision logic and systematically delivering precise, auditable actions. Modern methodologies support automated CGT induction from unstructured text, high-fidelity machine execution, and integration with AI agents for clinical decision support.

1. Formal Structure and Canonical Representation

A CGT is defined as a rooted, directed graph

$T = (V, E, r)$

where $V$ is the set of nodes, $E \subseteq V \times V$ is the set of directed edges, and $r \in V$ is the root node. Each node $v \in V$ is typed:

• Decision node: $\tau(v) = \textsc{Decision}$, carrying a clinical logical condition $C(v)$
• Outcome/Action node: $\tau(v) = \textsc{Outcome}$, storing a recommendation $O(v)$

Directed edges $(u \to v)$ carry branch labels such as “Yes”, “No”, “Mild”, “Severe” indicating the semantic outcome under a given condition. Trees are normalized post-extraction via standardized synonym rewriting (“elevated” → “>”), canonical units, uniform inequality formats, and the collapse of polytomous splits to binary when practical (Zhu et al., 2024).

For LLM-guided applications and computability, CGTs are serialized as JSON objects, bracketed S-expressions, or deterministic “if-elif-else” lists—all supporting recursive traversal, efficient matching, and robust audit (Li et al., 2023, Li et al., 16 May 2025, Deng et al., 7 Jan 2026).

2. Automated Extraction and Induction from Clinical Text

Modern CGT construction employs two principal paradigms (Zhu et al., 2024, Li et al., 6 Oct 2025):

• End-to-End LLM/Instruction-Tuned Generation: GPT-style autoregressive models are fine-tuned using (clinical text, serialized tree) pairs to directly emit CGT structures. Instruction tuning is driven by negative log-likelihood loss:

$V$0

Chain-of-thought prompting (“Let’s think step by step. …”) significantly improves structural accuracy, boosting F1 from 85.2% to 88.0% and tree-level accuracy from 58.2% to 61.4% (Zhu et al., 2024).

• Pipeline/Modular Extraction: CGT induction is split into three modules:
  1. Node Extraction: Span-level token classification (BERT/encoder) with BIO tagging labels all decision and outcome segments:

  $V$1

1. Edge Relation Classification: Pairwise edge prediction (e.g., via RoBERTa) classifies valid transitions (“Yes-edge”, “No-edge”, “None”).
2. Tree Assembly: A maximum spanning tree is assembled, enforcing acyclicity and canonical tree structure. Lightweight pipelines (BERT-Small/RoBERTa-small) with ~36M parameters achieve TreeAcc = 51.1%, offering high clinical deployment efficiency compared to 2B+ parameter generators (Zhu et al., 2024).

• Parameter-Efficient Fine-Tuning (PI-LoRA): LoRA-based adaptation with integrated gradient-path scoring allocates parameter budget to high-synergy modules, pruning and refitting to yield lean yet highly accurate CGT extractors. PI-LoRA demonstrates Tree_Acc=0.772, DP_F1=0.884, and Tree_LR=0.967—exceeding alternative PEFT strategies while tuning only ~0.4% of backbone weights (Li et al., 6 Oct 2025).
• Image-to-Tree (IEET) Extraction: Flowchart detection from guideline figures (via Faster R-CNN, OCR post-processing, and manual audit) scales CGT induction to cover >12 hospital departments with 1,202 trees, average depth 15–25 nodes. Flattened “If-Elif-Else” text representations support LLM execution (Li et al., 2023).

3. Tree Execution, LLM Integration, and Patient Interaction

CGTs are traversed deterministically or via LLM-guided agents, with each decision node evaluated against patient-specific context. Formally, for patient $V$2:

$V$3

where $V$4 tests the clinical predicate or feature at node $V$5, with operators over continuous and categorical variables (Oniani et al., 2024).

LLM-execution agents manage node queries as either explicit yes/no checks, multi-criteria satisfaction (count-of-true subpredicates compared to thresholds), or patient dialogue:

• For a condition node, LLM prompt: “Given the patient’s complaints $V$6 and dialog history $V$7, does the patient satisfy: ‘$V$8’? Answer Yes, No, or Unable to determine.”
• On “Unable to determine,” the agent autogenerates and queries for missing information, updating history and reevaluating the tree (Li et al., 2023, Deng et al., 7 Jan 2026).

This mechanism supports both batch scenario evaluation and real-time patient-LLM interaction. For scenario-based benchmarking, CGTs drive systematic MCQ instantiation covering combinatorial guideline branches (Lundin et al., 28 Aug 2025, Li et al., 16 May 2025).

4. Evaluation Metrics, Structural Fidelity, and Benchmarking

Quantitative evaluation is conducted at several abstraction levels:

• Node-level and Edge-level Metrics: Precision, recall, and F1:

$V$9

• Tree-level Accuracy: Correct reproduction of all nodes and edges:

$E \subseteq V \times V$0

• Structural Edit Distance: Mean normalized tree-edit distance between prediction and gold standard.
• Clinical Relevance Benchmarking: Scenario-based MCQs and guideline-concordant decision chains are used to test LLMs (e.g., MedGUIDE, ~7,747 expert-validated MCQs, mean accuracy 0.64 for GPT-4.1, 0.25 for domain LLMs) (Li et al., 16 May 2025). Graph-based test-harness systems scale guideline coverage to trillions of scenario combinations, driving granular error analysis and LLM post-training (Lundin et al., 28 Aug 2025).

CoT prompting, explicit guideline JSON inclusion, and continued pretraining all yield measurable gains in guideline adherence and protocol-consistent recommendations (Zhu et al., 2024, Li et al., 16 May 2025, Deng et al., 7 Jan 2026).

Experimental results from Table 1 (Text2MDT) summarize pipeline comparison (Zhu et al., 2024):

Method	Node F1	Edge F1	Span-Rel F1	TreeAcc
End2End GPT (no CoT)	78.5	82.1	80.0	36.4
End2End GPT (+CoT)	81.0	85.4	83.1	61.4
Pipeline (RoBERTa-Large)	89.8	94.6	92.2	64.8
Light Encoder Pipeline	83.3	93.6	88.3	51.1

5. Clinical Decision Support, Visualization, and Human Interaction

CGTs are the backbone for clinical decision support systems (CDSS), directly enabling evidence-based, auditable decision workflows:

• Interactive Navigation: Multi-path CGTs visualized with fisheye layouts as in Orient-COVID facilitate one-screen navigation of complex protocols, supporting user actions (select, answer, backtrack) and ensuring only guideline-valid branches are enabled (Jammal et al., 2024).
• Adherence Impact: Randomized simulation trials demonstrate that CDSS with CGT-based navigation significantly improve adherence (mean total score 17.02 vs. 15.42, $E \subseteq V \times V$1), especially for critical management actions (e.g., troponin testing 31%→57%, anticoagulant prescription 70%→98%) (Jammal et al., 2024).
• Error Handling and Runtime Monitoring: Behavior-Tree analogs to CGTs support composable subtrees, parallel and sequence constructs, runtime state monitoring, and decorator nodes for recovery, with real-time error detection and logging crucial for high-risk intervention protocols (Hannaford et al., 2018).

CGTs are also used in subgroup identification, where recursively partitioned trees define patient subgroups with homogeneous treatment effects, and covariate-adjusted estimators inform split and pruning criteria (Steingrimsson et al., 2018).

6. Adaptation, Generalization, and Practical Recommendations

Translating extraction and deployment pipelines across clinical domains and languages requires careful lexical normalization (units, synonyms), handling of negations, and adaptation to different text structures (tables, bullets, coreferences) (Zhu et al., 2024, Deng et al., 7 Jan 2026).

Key recommendations:

• Incorporate biomedical pretraining (e.g., PubMedBERT) for node extraction modules.
• Extend normalization dictionaries for unit/categorical harmonization in English protocols.
• Leverage weak supervision from structured sections (e.g., “Recommendation #1”) and investigate joint multi-task learning (joint node/edge extraction) to reduce error propagation.
• Implement human-in-the-loop authoring and validation UIs to expedite clinical acceptance and protocol fidelity (Zhu et al., 2024).

For high-stakes, high-fidelity use, manually curated or validated CGTs are scaffolded to ensure exact logic, with LLMs executing fixed traversal protocols, eliminating “hallucinated” inferences (Oniani et al., 2024, Deng et al., 7 Jan 2026).

7. Benchmarking, Safety, and Limitations

Even the best domain-adapted LLMs show significant deficits in strict guideline adherence and tree-based reasoning without explicit CGT context. Inclusion of tree structure in context and pretraining on protocol data are mandatory for robust, safe clinical operation. Robust audit trails, rationale logging, and reward-model-based deviation detection are recommended for production deployments (Li et al., 16 May 2025).

The CGT formalism, validated through large-scale dataset construction and multi-domain evaluation, remains the preeminent computational abstraction for clinical protocol execution, interpretation, and AI-enabled decision support.