Source-linked AI summary
Search Broadly, Seek Evidence on Both Sides, Decide Narrowly: Evidence-Admissible GraphRAG for Longitudinal Clinical Event Verification
Xingtao Lin, Yubo Feng, Weixin Liu, Hangqi Ren, Junchao Zhou, Caiwan Sun, You Chen
TL;DR
Longitudinal clinical event-relation verification must recover linked occurrences, distinguish supporting from contradicting evidence, and ensure that conclusions use patient-consistent, source-resolvable evidence. MedEventGraph-RAG separates broad discovery from narrow evidence-admissible assessment, improving verification across tasks while reducing unsupported conclusions.
Problem
Longitudinal records distribute evidence across observations, notes, encounters, and time, while negation, temporal mismatch, redundancy, and conflicting findings can make relevant information insufficient to establish a relation.
Method
MedEventGraph-RAG uses an event-centric patient graph and source-linked retrieval to discover candidate chains and a query-specific evidence contract to admit only patient-consistent, occurrence-bound, scope-valid, source-resolvable evidence for separate support and refutation assessment.
Results
Across verification tasks, B6 improves balanced accuracy over the strongest matched control by 26.9, 4.9, and 30.4 points, reaches 92.2 balanced accuracy with zero false support under masked evidence, and retains 57.9 provenance-valid reconstructed chains.
Takeaways & Limitations
Separating discovery from assessment keeps retrieval signals from determining conclusions, supports broad evidence search without rejecting legitimate cross-encounter evidence, and localizes the remaining gap to source binding.
Takeaways & Limitations
Evidence isolation is materially more expensive than the lower-bound configuration, requiring 4.625 calls and 12.8k tokens per query versus 1.000 calls and 1.4k tokens.
Abstract
from arXiv · showhide
Longitudinal clinical event-relation verification determines whether a patient record supports a specified relation among two or more clinical events. This task is challenging because evidence is distributed across structured records, notes, laboratory trajectories, encounters, and time, while negation, temporal mismatch, repeated documentation, and conflicting findings can make retrieved information appear relevant without establishing the relation. We present MedEventGraph-RAG, an evidence-admissible framework that represents event occurrences in a patient-specific graph and links each occurrence to source evidence, including structured rows, note spans, timestamps, and numerical trajectories. Given a verification query specifying events, relation, and clinical scope, the graph guides discovery of candidate event chains and retrieves evidence from both supporting and contradicting sides. A query-specific evidence contract filters information by patient identity, scope, occurrence binding, and source traceability before a separate assessor determines supported, conflicting, refuted, or insufficient outcomes. Across ten protocols on i2b2, n2c2, MIMIC-IV, and LUNGUAGE, MedEventGraph-RAG achieves balanced accuracies of 78.6, 67.3, and 96.8 on temporal, medication-adverse-event, and recorded-order verification, improving over the strongest matched baselines by 26.9, 4.9, and 30.4 points. Under evidence masking, it reaches 92.2 balanced accuracy with no false-support predictions. When intermediate events are hidden, it recovers complete source-traceable event chains in 57.9% of i2b2 and 70.0% of LUNGUAGE cases. These results show that separating broad evidence discovery from narrow evidence-admissible assessment improves longitudinal clinical verification and reduces unsupported conclusions.
Introduction
Longitudinal verification must recover complete event chains, assess support and contradiction, and ground conclusions in traceable patient evidence. MedEventGraph-RAG separates broad two-sided discovery from narrow assessment using validated evidence.
- Motivation: Clinical relations are implicit across structured records, notes, encounters, laboratory trajectories, and time.Relevant chains may connect medication administration, a later creatinine rise, and a diagnostic assessment.
- Motivation: Negation, copied documentation, temporal mismatch, and conflicting findings can make relevant observations non-evidential.These issues can also mix incompatible episodes or substitute documentation time for occurrence time.
- Requirements: Reliable verification requires recovering event occurrences, determining whether evidence supports or contradicts the relation, and grounding every event.The framework targets these capabilities jointly across longitudinal records.
- Proposed framework: MedEventGraph-RAG searches broadly across records, seeks supporting and contradictory observations, and decides only from validated patient evidence.An event-centric graph and stateful controller discover candidate chains, while a separate assessor evaluates support and refutation.
- Proposed framework: A query-specific evidence contract admits only patient-consistent, scope-valid, and source-resolvable evidence before assessment.Retrieval signals can guide recovery but cannot determine the conclusion.
Related Work
Prior work broadens retrieval through graph, adaptive, clinical, and memory-based methods, but patient-specific memory can introduce cross-record leakage. MedEventGraph-RAG therefore retains only patient-independent retrieval policies.
- Graph and adaptive RAG: Graph-based and adaptive RAG methods broaden candidate discovery through relational, graph–vector, textual-subgraph, and iterative graph–document retrieval.Clinical variants add medical knowledge, patient history, temporal structure, and factual or counterfactual context.
- Agent memory: Agent-memory methods reuse reflections or workflows to improve future behavior.These approaches motivate procedural reuse but raise record-isolation concerns in patient-specific verification.
- Agent memory: MedEventGraph-RAG retains only patient-independent retrieval policies to avoid an unintended cross-record channel.Its memory contains no patient identifiers or facts.
Method
The method separates graph-guided, two-sided evidence discovery from frozen-bundle assessment. It binds occurrences to source contexts, filters evidence through a contract, and verifies each relation in compositional chains.
- Query and representation: The query is normalized into event slots, relation constraints, and encounter-, episode-, or patient-level scope.The system returns supported, conflicting, refuted, or insufficient with source-resolvable evidence and citations.
- Algorithm: Algorithm 1 iterates broad expansion, two-sided retrieval, binding and deduplication, admissibility filtering, assessment, and retrieval replanning.Replanning changes retrieval only; assessment remains isolated from policy, rank, and memory.
- Query and representation: Patient event graphs index occurrences and source-linked contexts, while graph structure guides discovery rather than establishing relations.Contexts preserve timestamps, negation, uncertainty, historical status, values, and sources.
- Broad two-sided discovery: Adaptive discovery expands patient-scoped event chains and retrieves aligned and counteroriented observations across encounters and time.Temporally or typologically incompatible patterns remain eligible because they may contain counter-evidence.
- Broad two-sided discovery: Discovery signals rank candidates, but weak signals lower priority without acting as hard filters.Transition-evidence coverage measures whether each relation can be examined, not whether it supports the query.
- Evidence contract: The frozen evidence bundle contains occurrence-bound contexts, recorded time information, and provenance, while retrieval metadata remains outside assessment.Only patient-consistent, scope-valid, occurrence-bound, source-resolvable items are admitted.
- Narrow assessment: The assessor evaluates temporal compatibility and independent support and refutation axes from the admissible bundle.Evidence on both sides raises both axes, while unresolved evidence raises neither.
- Compositional verification: A compositional query is supported only when every occurrence and adjacent relation has scope-consistent, admissible evidence.Recovering graph nodes or order alone is insufficient; hidden-intermediate recovery separates structural reconstruction from evidential verification.
Task Construction
The study constructs ten protocols spanning clinical relation verification, retrieval, masking, transfer, joint-chain assessment, and hidden-intermediate recovery, with controlled sources, labels, negatives, and artifacts.
- Query sources and labels: Pairwise queries use i2b2 TLINK, n2c2 medication–ADE, and MIMIC-IV recorded-order annotations, with labels held out for evaluation.Relation direction is excluded from events, edges, contexts, prompts, retrieval scores, and memory.
- Negatives: Negatives perturb the relation of a positive chain while preserving its events and contexts, preventing resolution through superficial structural differences.On T9 i2b2, positive and negative chains have similar lengths, hops, and context counts; T9 n2c2 classes are identical on these measures.
- The ten protocols: Ten protocols cover pairwise relation verification, retrieval alone, evidence masking, procedural-memory effects, transfer, joint relations, and hidden-intermediate chain recovery.T1–T3 use i2b2, n2c2, and MIMIC-IV; T10 additionally uses LUNGUAGE.
- What is held fixed: All methods receive identical frozen artifacts, identifiers, patient filters, encoders, candidate caps, context budgets, and assessors; only the online retrieval operator varies.No protocol is filtered by any system’s output.
Experiments
Across controlled protocols, MedEventGraph-RAG improves verification accuracy and evidence integrity, while compositional evaluation localizes remaining losses to provenance binding. These gains require substantially higher inference cost.
- Benchmark results: 26.9, 4.9, and 30.4 BA gains over the strongest matched controls establish improvements across the three pairwise tasks.On the matched index, MedEventGraph-RAG reaches 78.6, 67.3, and 96.8 BA versus external best results of 72.3, 61.9, and 78.8.
- Evidence integrity: 75.5/75.0 on T6 preserves 76.3% supported recall while improving over B5 by 23.6 BA, showing that admissibility need not discard legitimate cross-encounter evidence.The ablation attributes gains separately to finding evidence and judging it, while verifier-guided replanning further raises chain recall to 55.0.
- Compositional verification: 99.9 and 77.1 through verification fall to 57.9 provenance-valid recall, identifying source binding as the compositional bottleneck.Richer contexts raise provenance to 70.0 at unchanged C@20, indicating a contextual rather than architectural limitation.
- Procedural memory: B7 improves T6 by 1.2/1.3 BA/F1 at unchanged false-support rate, adding retrieval benefit without changing judgment.The memory stores no patient facts, spans, answers, or verdicts and equals B6 on T1–T3 and T5.
- Cost boundary: Evidence isolation costs 4.6× calls and 9.3× tokens per query compared with B2, while procedural memory adds 3.06 seconds of local work.B6 uses 4.625 calls and 12.8k tokens per query versus 1.000 and 1.4k for B2.
Discussion and Conclusion
The paper separates discovery from judgment to address three independent failure modes in longitudinal verification, while limiting its claims to tested corpora and avoiding causal or clinical-safety conclusions.
- Event-occurrence and source binding raises hidden-bridge recovery to 99.9.
- Independent support and refutation axes produce zero false support under masked evidence.
- Assessment-time evidence contracts preserve the patient boundary under distractors while retaining 76.3% supported recall.