Framework for Grounding Healthcare LLMs in a Causal Knowledge Graph: A Cardiovascular Example Pilot

Ummara Mumtaz; Aimen Noor; Awais Ahmed

cs.AIAug 15, 2026

Framework for Grounding Healthcare LLMs in a Causal Knowledge Graph: A Cardiovascular Example Pilot

Authors: Ummara Mumtaz, Aimen Noor, Awais Ahmed

Abstract

Large language models (LLMs) are increasingly proposed for healthcare decision support, but their evaluations still reward single-answer accuracy rather than reasoning about interventions, mechanisms, harms, evidence, and uncertainty. We propose a reproducible, graph-centered evaluation framework for intervention-oriented LLM behavior in healthcare and stress-test it in a cardiovascular pilot. The framework has four components: (i) a domain causal knowledge graph in which assertions are first-class, provenance-preserving nodes with stable identifiers; (ii) a scenario-conditioned subgraph extraction step that, given any clinical scenario, retrieves the relevant reified-assertion subgraph; (iii) four controlled grounding conditions that vary how the retrieved subgraph is composed into the model's context (ungrounded C1, knowledge-graph C2, causal-graph C3, integrated C4); and (iv) an automated scoring pipeline, anchored on assertion identifiers, that computes intervention accuracy, and other evaluation measures on a single pass. To test the framework, we built a category-balanced scenario generator across eight reasoning failure modes and instantiated it on a cardiovascular graph. The metric panel discriminates conditions along interpretable, non-redundant axes: C4 obtains the strongest causal edge F1 (0.838), adverse-effect F1 (0.833), evidence accuracy (0.738), and unsupported claim rate (0.114), while C1 obtains the highest raw intervention accuracy (0.948) with no measurable causal or evidential grounding.

Figures & tables

Figure 1: Framework pipeline from reified-assertion knowledge graph to scored model output, showing the four coupled components: knowledge graph, scenario-conditioned subgraph extraction, controlled grounding conditions (C1–C4), and automated scoring anchored to assertion_id s.

Figure 2: Partial visualization of the constructed Neo4j biomedical knowledge graph, showing selected biomedical entities and their interconnected relationships.

Figure 3: Assertion reification schema. (a) An edge-only representation of a biomedical claim, in which provenance, confidence, population, and validation cannot be attached flexibly. (b) Reified representation: the claim is a first-class Assertion node with a stable identifier, linked to its subject and object through HAS_SUBJECT and HAS_OBJECT , and to one or more Evidence Source nodes through SUPPORTS .

Figure 4: Worked example of a gold causal pathway. Each biomedical entity is connected to the next through a reified assertion node with a stable identifier, so that a gold pathway is expressible as an ordered list of assertion_id s and consumed directly by the scoring metrics for causal edge precision, recall, and F1.

Stage	Content
Scenario input	62-year-old with essential hypertension, LDL 138 mg/dL, no diabetes, eGFR 78, not pregnant. Decision target: choose the preferred first-line antihypertensive from {ACE inhibitor, thiazide diuretic, beta-blocker, calcium-channel blocker}.
Gold intervention	ACE inhibitor (lisinopril).
Gold causal pathway	[ASR_0019, ASR_0010, ASR_0011, ASR_0013] $=$ ACE inhibitor TARGETS ACE $\rightarrow$ PARTICIPATES_IN RAAS activity $\rightarrow$ CAUSES_DECREASE_IN blood pressure $\rightarrow$ CAUSES_DECREASE_IN cardiovascular event risk.
Gold adverse effects	{ ASR_0087 (dry cough), ASR_0091 (hyperkalemia)}.
Gold contraindication scope	Nonapplicable in this scenario (patient is not pregnant); metric is undefined.
Gold evidence sources	{ EV_0007 (JNC-8 guideline), EV_0011 (HOPE trial)}.

Table 1: Worked example of one scenario passing through the scoring pipeline. The scenario is drawn from the direct-effect category; identifiers ( ASR_* , EV_* ) refer to nodes in the reified-assertion graph. The example is illustrative and does not affect the aggregate numbers in Section 4 .

Metric	C1 (mean, $n$ )	C2 (mean, $n$ )	C3 (mean, $n$ )	C4 (mean, $n$ )
Intervention accuracy	0.948, 77	0.873, 79	0.808, 78	0.886, 79
Preferred intervention accuracy	0.948, 77	0.873, 79	0.808, 78	0.886, 79
Causal edge precision	NA, 0	0.522, 79	0.808, 78	0.806, 79
Causal edge recall	0.000, 77	0.785, 79	0.808, 78	0.886, 79
Causal edge F1	NA, 0	0.579, 79	0.808, 78	0.838, 79
Adverse-effect F1	NA, 0	0.733, 20	0.822, 30	0.833, 40

Table 2: Overall metric panel across grounding conditions in the cardiovascular pilot demonstration (gpt-5.4). NA indicates the metric is undefined for that condition given the scoring rule. Lower is better for unsupported claim rate; higher is better for all other metrics.

Metric	C2 mean	C4 mean	$\Delta$ (C4 $-$ C2)	Direction favorable to C4?
Intervention accuracy	0.873	0.886	$+0.013$	Yes
Causal edge precision	0.522	0.806	$+0.284$	Yes
Causal edge recall	0.785	0.886	$+0.101$	Yes
Causal edge F1	0.579	0.838	$+0.259$	Yes
Adverse-effect F1	0.733	0.833	$+0.100$	Yes
Evidence accuracy	0.705	0.738	$+0.033$	Yes

Table 3: Illustrative within-scenario paired deltas for the C4 vs. C2 contrast in the pilot. Positive deltas indicate improvement for C4 on the metric; for unsupported claim rate, a negative delta is favorable.

LLM Grounding

Large language model (LLM) grounding studies how models anchor answers and actions in evidence, structured knowledge, and real-world context.

See all 319 papers

Knowledge-Graph Grounding Helps LLMs Only for Out-of-Training Knowledge: A Controlled Study on Clinical Question Answering
Jun 21, 2026Madhulatha Mandarapu, Sandeep KunkunuruFactual Knowledge in Language Models LLM Reasoning with Graphs
samyama
Why LLMs Fail at Causal Discovery and How Interventional Agents Escape
May 26, 2026Amartya Roy, Sonali ParbhooCausal Discovery Causal Reasoning in Language Models
SIRE, IIT Delhi and Robert Bosch GmbH, India · Imperial College London
Clinical Reasoning Graphs: Structured Evaluation of LLM Diagnostic Reasoning Reveals Competence Without Consistency
Jun 29, 2026Nisarg A. PatelLLM Evaluation Reasoning Consistency in Language Models
University of California, San Francisco · Department of Oral and Maxillofacial Surgery, University of California, San Francisco, CA, USA · Department of Medicine, Division of Clinical Informatics and Digital Transformation, University of California, San Francisco, CA, USA

Jun 21, 2026J/K move · Enter open · S save

Knowledge-Graph Grounding Helps LLMs Only for Out-of-Training Knowledge: A Controlled Study on Clinical Question Answering

Madhulatha Mandarapu, Sandeep Kunkunuru

samyama

Factual Knowledge in Language Models LLM Reasoning with Graphs Clinical QA LLM Grounding Clinical Language Model Evaluation Biomedical KGs+1

Open paper AskPDF arXiv

A recent Nature Medicine study reports that general-purpose frontier LLMs outperform specialized retrieval-augmented clinical tools on medical benchmarks, and that retrieval can hurt strong models. We ask the natural follow-up: does structured knowledge-graph (KG) grounding change this, and when does grounding help at all? We contribute two results. First, a reproduction: the study's headline HealthBench score (~88) is the Consensus variant, not full HealthBench, where frontier models and ideal completions both score ~46-47 under a physician-calibrated grader (agreement 82.5%); we reproduce GPT-5.2 Consensus =90.9 and flag a score-deflating grader bug. Second, a knowledge-boundary result. Using a graph+vector engine (samyama-graph) over the public biomedical KG PrimeKG, neither naive triple retrieval nor an agentic natural-language-to-Cypher loop (82% successful queries) improves MedQA across a weak-to-strong model ladder (all |Delta| <= 3.4). On a synthetic counterfactual KG, and on a hybrid benchmark mixing known and novel facts, the identical pipeline lifts out-of-training accuracy from chance to ~100% (+68 to +79) while adding nothing on known facts (a no-LLM arm answers both). Across three regimes (no-knowledge, graph-aided, hybrid), grounding helps only insofar as the decisive fact lies outside the model's training -- public-KG facts are redundant, private and novel data are where it pays -- matching the study's institutional-data caveat.

Abstract

LLM Grounding

Explore similar work

Knowledge-Graph Grounding Helps LLMs Only for Out-of-Training Knowledge: A Controlled Study on Clinical Question Answering

Why LLMs Fail at Causal Discovery and How Interventional Agents Escape

Clinical Reasoning Graphs: Structured Evaluation of LLM Diagnostic Reasoning Reveals Competence Without Consistency