Paper deep dive
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
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 8/25/2026, 8:03:12 AM
Summary
The paper introduces MedEventGraph-RAG, an evidence-admissible framework for longitudinal clinical event verification. It constructs patient-specific graphs linking event occurrences to source evidence (structured rows, notes, trajectories) and employs a two-sided retrieval strategy to find supporting and contradicting evidence. A strict evidence contract filters this information based on patient identity, scope, and traceability before a separate assessor determines if a relation is supported, conflicting, refuted, or insufficient. The method achieves significant improvements in balanced accuracy over baselines on datasets like i2b2, n2c2, MIMIC-IV, and LUNGUAGE.
Entities (10)
Relation Signals (9)
MedEventGraph-RAG ā evaluatedon ā MIMIC-IV
confidence 98% Ā· Across ten protocols on i2b2, n2c2, MIMIC-IV, and LUNGUAGE, MedEventGraph-RAG achieves balanced accuracies
MedEventGraph-RAG ā evaluatedon ā LUNGUAGE
confidence 98% Ā· Across ten protocols on i2b2, n2c2, MIMIC-IV, and LUNGUAGE, MedEventGraph-RAG achieves balanced accuracies
MedEventGraph-RAG ā evaluatedon ā i2b2
confidence 98% Ā· Across ten protocols on i2b2, n2c2, MIMIC-IV, and LUNGUAGE, MedEventGraph-RAG achieves balanced accuracies
MedEventGraph-RAG ā evaluatedon ā n2c2
confidence 98% Ā· Across ten protocols on i2b2, n2c2, MIMIC-IV, and LUNGUAGE, MedEventGraph-RAG achieves balanced accuracies
MedEventGraph-RAG ā solves ā Longitudinal Clinical Event Verification
confidence 97% Ā· We present MedEventGraph-RAG, an evidence-admissible framework ... for Longitudinal Clinical Event Verification
MedEventGraph-RAG ā uses ā Patient Event Graph
confidence 95% Ā· MedEventGraph-RAG ... represents individual event occurrences in a patient-specific graph
MedEventGraph-RAG ā applies ā Evidence Contract
confidence 94% Ā· A query-specific evidence contract filters information ... before a separate assessor determines supported, conflicting, refuted, or insufficient outcomes.
ā ā
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2608.22062v1
- Canonical: https://arxiv.org/abs/2608.22062v1
Trouble viewing inline? Open PDF directly ā
Full Text
45,522 characters extracted from source content.
Expand or collapse full text
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 Abstract Longitudinal clinical event-relation verification determines whether a patient record supports a specified relation among two or more clinical events, such as whether acute kidney injury was documented after vancomycin within one encounter. Verification is difficult because the relevant information may be distributed across structured records, notes, laboratory trajectories, encounters, and time, while repeated documentation, negation, temporal mismatch, and conflicting findings can make retrieved information appear relevant without establishing the relation. We present MedEventGraph-RAG, which represents individual event occurrences in a patient-specific graph and links each to its record source: structured rows, note spans, timestamps, local context, and numerical trajectories. A verification query specifies the events, the relation, and the clinical scope to be evaluated. The graph then guides retrieval of candidate event sequences connecting those events, together with source evidence that may support or contradict the relation, and redirects retrieval while required events or relations remain unresolved. Before assessment, the system excludes information belonging to another patient, outside the specified encounter-, episode-, or patient-level scope, or lacking a traceable source. A separate assessor evaluates only the resulting query-specific evidence bundle and returns supported, conflicting, refuted, or insufficient; graph scores, retrieval rankings, search history, and procedural memory guide retrieval but do not enter that bundle. Across ten protocols on i2b2, n2c2, MIMIC-IV, and LUNGUAGE, it reaches balanced accuracies of 78.6, 67.3, and 96.8 on pairwise temporal, medicationāadverse-event, and recorded-order verification, exceeding the strongest matched baseline by 26.9, 4.9, and 30.4. When a required evidence component is removed it reaches 92.2 balanced accuracy with no false-support predictions; when intermediate events are withheld it returns complete event sequences with traceable evidence for every relation in 57.9% of i2b2 and 70.0% of LUNGUAGE cases. These results indicate that pairing patient event graphs with source-linked evidence can improve longitudinal verification and reduce unfounded conclusions when evidence is missing. Introduction Longitudinal clinical event-relation verification asks whether a patientās electronic health record supports a given statement about the relation between two or more clinical events. This task is challenging because electronic health records represent medications, laboratory measurements, diagnoses, procedures, and clinical assessments as separate observations across structured tables, narrative notes, encounters, and time, leaving clinically meaningful relations largely implicit (28; 2). The required evidence may therefore form a chain linking medication administration, a subsequent creatinine rise, and a diagnostic assessment, while apparently relevant observations may be negated, copied forward, temporally misaligned, or drawn from another episode. Reliable verification consequently requires three capabilities: recovering the event occurrences needed to complete the chain, determining whether the recovered evidence supports or contradicts the specified relation, and grounding every event and relation in the correct patient, clinical scope, timestamp, and source. These three capabilities fail for distinct reasons. Copied notes, templated sections, and recurrent event types produce redundant mentions that exhaust a limited retrieval budget while the occurrences needed to ground a cross-encounter relation go unretrieved (32). Relevance does not establish direction, because retrieved content may affirm, negate, qualify, or retrospectively mention an event, and correctly identified events may still be temporally incompatible with the queried relation or contradicted elsewhere (4; 10; 30). Admissibility fails last: a plausible chain may combine incompatible episodes, substitute documentation time for occurrence time, or rest on copied statements whose originating evidence cannot be resolved. Verification therefore needs broad signals to discover candidate chains, but stricter criteria to decide from them. We propose MedEventGraph-RAG, an evidence-admissible framework that searches broadly across longitudinal records, seeks evidence on both sides of a verification query, and decides only from validated patient evidence. An event-centric patient graph and a stateful controller discover query-covering candidate chains across structured records, notes, numerical trajectories, encounters, and time, linking every recovered occurrence to its originating row, span, or trajectory and redirecting search while required events, transitions, or contexts remain unresolved. The controller retrieves relation-aligned and potentially contradictory observations alike, and a separate assessor then evaluates support and refutation on independent axes, distinguishing evidence on both sides from evidence on neither. A query-specific evidence contract finally admits only patient-consistent, scope-valid, and source-resolvable evidence, so graph scores, retrieval priors, controller state, and procedural memory can influence what is recovered but never what is concluded. Figure 1: MedEventGraph-RAG in one view. Existing RAG paradigms may retrieve relevant but non-evidential content, whereas our framework binds event paths to source contexts and isolates assessment from search signals. Related Work Graph-based and adaptive RAG broaden candidate discovery through relational expansion, graphāvector fusion, textual-subgraph retrieval, and iterative graphādocument search (37; 8; 9; 7; 6; 25; 13; 20; 18; 11; 19; 23; 1; 34). Clinical variants incorporate medical knowledge, patient history, temporal structure, and factual or counterfactual context for generation, prediction, and longitudinal retrieval (33; 36; 14; 3; 27). These methods improve semantic access and structural reach, but do not separate the signals used to retrieve candidates from the evidence allowed to determine a patient-specific verdict. Domain-adapted clinical LLMs improve medical language understanding without enforcing this boundary (5; 17). Agent-memory methods reuse reflections or workflows to improve future behavior (26; 35; 31; 22), but instance-level memory could create an unintended cross-record channel in patient-specific verification. We instead retain only patient-independent retrieval policies. Method List of algorithms 1 Broad, two-sided discovery with evidence-admissible assessment. Input: query q; patient graph pG_p; context store pC_p; optional memory ā³M; budgets K,B,dmax,RmaxK,B,d_ ,R_ Output: verdict y; evidence axes (S+,Sā)(S^+,S^-); frozen bundle ā¬cB_c 1: cāParseā(q)cā Parse(q); uāInitObligationsā(c)uā InitObligations(c) 2: ĻāInitPolicyā(c,ā³)Ļā InitPolicy(c,M); rā0rā 0 3: F0āAnchorsā(c,p)F_0ā Anchors(c,G_p); āāā Iā 4: for d=1d=1 to dmaxd_ do ā³ search broadly, on both sides 5: PdāExpandā(Fdā1,p,K,Ļ)P_dā Expand(F_d-1,G_p;K,Ļ) 6: (Cd+,Cdā)āTwoSidedRetrieveā(Pd,c,u,p)(C_d^+,C_d^-)ā TwoSidedRetrieve(P_d,c,u;C_p) 7: (Pd,Cd)āBindDedupā(Pd,Cd+āŖCdā)(P_d,C_d)ā BindDedup(P_d,C_d^+āŖ C_d^-) 8: (Pd,u)āCandidateCheckā(Pd,Cd,c,u)(P_d,u)ā CandidateCheck(P_d,C_d,c,u) 9: P^dāTopB(Pd;R(ā ā£c),B) P_dā TopB(P_d;R(Ā· c),B) ā³ decide narrowly, from admissible evidence only 10: for each PāP^dPā P_d do 11: āāSourceDedupā(āāŖContractā(c,P,Cdā(P)))Iā SourceDedup (IāŖ Contract(c,P,C_d(P)) ) 12: ā¬cā(c,ā)B_cā(c,I) ā³ excludes Ļ,R,ā³Ļ,R,M 13: (S+,Sā,y)āAssessĪøā(ā¬c)(S^+,S^-,y)ā Assess_Īø(B_c) 14: uāUpdateObligationsā(u,c,ā¬c)uā UpdateObligations(u,c,B_c) 15: if y=conflictingy=conflicting or Resolvedā(u,c) Resolved(u,c) then break ā³ replan retrieval only, never assessment 16: fāStructuredFailureā(u,c)fā StructuredFailure(u,c) 17: if fā ā fā and r<Rmaxr<R_ then 18: ĻāReplanā(Ļ,f,ā³)Ļā Replan(Ļ,f,M); rār+1rā r+1 19: FdāFrontierā(Pd,u,Ļ)F_dā Frontier(P_d,u,Ļ) 20: return y, (S+,Sā)(S^+,S^-), ā¬cB_c Figure 2: Patient event graphs index source-resolvable contexts; adaptive two-sided discovery builds a query-specific bundle, and only admissible evidence reaches the fixed-bundle assessor. Task and information-flow invariants. Given a record p and query q, we parse a normalized verification query c=(E,R,Ļ)c=(E,R,Ļ), where E holds observed and unresolved event slots, R their relation constraints, and Ļāencounter,episode,patientĻā\encounter,episode,patient\ the clinical scope. Unspecified fields stay unresolved. The system returns yāsupported,conflicting,refuted,insufficientyā\supported,conflicting,refuted,insufficient\ with its source-resolvable evidence and citations. The information flow is Dc D_c =Discoverā(c,p,p,Ļ), = Discover(c,G_p,C_p;Ļ), ā¬c _c =Contractā(c,Dc), = Contract(c,D_c), (S+,Sā,y) (S^+,S^-,y) =Assessā(c,ā¬c), = Assess(c,B_c), (1) where pG_p and pC_p are the patient event graph and source-linked context store, Ļ is the retrieval policy, DcD_c includes candidates and search metadata, and ā¬cB_c contains only admitted evidence. Connectivity and retrieval relevance nominate candidates but do not establish the relation, and the direction in which evidence was retrieved does not fix its evidential polarity. Every item in ā¬cB_c must be patient-consistent, occurrence-bound, source-resolvable, and not scope-invalid, while graph scores, retrieval ranks, controller state, search history, and memory remain outside ā¬cB_c and cannot alter the assessment of an identical bundle. Search broadly: patient eventācontext representation. Each record is (p,p)(G_p,C_p), separating structural navigation from evidential interpretation. pG_p indexes occurrences with their recorded encounter and temporal structure; its edges are limited to recorded relations and deterministic derivations, never gold relations, assessor judgments, or model-inferred relations. pC_p maps each occurrence to its originating structured row, narrative span, or numerical trajectory, preserving negation, uncertainty, historical status, values, timestamps, and sources. Thus pG_p broadens discovery, while the source-linked contexts supply the evidence from which relations are actually assessed. Search broadly and seek evidence on both sides: adaptive discovery. Given the normalized query c, discovery initializes a patient-scoped frontier and iteratively expands candidate event chains across time and encounters. At depth d the controller traverses structurally valid neighbors without requiring agreement with the specified relation, retrieving their occurrence-linked contexts through lexical, dense, and metadata-aware channels. Retrieval maintains complementary obligations: aligned queries seek observations consistent with the specified events and relations, while counter-oriented queries target negation, incompatible order, historical or hypothetical mentions, stable findings, and conflicting documentation. Retrieved contexts are bound to candidate occurrences, deduplicated, and annotated with patient, scope, timestamp, and source status. Malformed and cross-patient candidates are removed, while temporal or type patterns incompatible with the specified relation stay eligible, because they are exactly where counter-evidence tends to appear. Because only a bounded number of candidates can be contracted, each path P receives a query-conditioned priority score Rā”(Pā£c)=ākāā”(c)[Ļµ+(1āĻµ)āskā(P,c)],R(P c)= _k (c) [Īµ+(1-Īµ)s_k(P,c) ], (2) where Ļµ=0.2Īµ=0.2 and each skā(P,c)ā[0,1]s_k(P,c)ā[0,1] is a normalized discovery signal measuring graph diffusion, path compactness, source-context relevance, source-link completeness, or temporal and event-type compatibility. For a compositional query c=(e0,r1,e1,ā¦,rh,eh)c=(e_0,r_1,e_1,ā¦,r_h,e_h), ā”(c)K(c) additionally includes transition-evidence coverage scov(P,c)=1hāi=1h[ĪP(eiā1,ri,ei)ā ā ],s_cov(P,c)= 1h _i=1^hI\! [ _P(e_i-1,r_i,e_i)ā ], (3) where ĪPā(eiā1,ri,ei) _P(e_i-1,r_i,e_i) denotes the source-linked contexts, timestamps, or trajectories available to assess transition (eiā1,ri,ei)(e_i-1,r_i,e_i) on P. This measures whether each transition can be examined, not whether it supports the specified relation, so disagreement does not by itself eliminate a path. The lifted product maps every component to [Ļµ,1][Īµ,1], so weak signals lower priority without acting as hard filters. At each depth the highest-priority candidates are contracted and assessed after source-level deduplication; path rank and depth are not aggregation weights. If an obligation remains unresolved the controller revises its policy from the corresponding structured failure and continues, stopping when both obligations resolve, admissible evidence establishes conflict, or the budget is exhausted. In Algorithm 1, dmaxd_ bounds event-to-event traversal and RmaxR_ bounds policy revisions. Decide narrowly: evidence contract. Discovery output is converted into a frozen, query-specific evidence bundle ā¬c=(c,āc),āc=zj=(Ej,Cj,Tj,Ī j)j=1m,B_c=(c,I_c), _c=\z_j=(E_j,C_j,T_j, _j)\_j=1^m, (4) where each item zjz_j binds one or more occurrences EjE_j to their source contexts CjC_j, recorded time types and granularities TjT_j, and structured-row or note-span provenance Ī j _j, keeping contexts, timestamps, and citations attached to what they ground. An item z is admitted only if Admā”(z,c) (z;c) =Apatientā(z)ā§Abindingā(z)ā§Asourceā(z) =A_patient(z) A_binding(z) A_source(z) ā§[Ascope(z;c)ā invalid], [A_scope(z;c) ], (5) where Ascopeā(z,c)āvalid,unknown,invalidA_scope(z;c)ā\valid,unknown,invalid\. Every admitted item must therefore resolve to the target patient, to specific occurrences, and to original sources, and items used jointly must be consistent with the query-defined scope. Missing episode metadata or coarse timing stays explicitly unknown: retained for assessment, but unable by itself to establish support or refutation. The contract enforces patient identity, scope consistency, occurrenceāsource binding, and source resolution at assessment time; it does not guarantee correct event extraction, clinical interpretation, or the factual accuracy of the underlying record. Decide narrowly: isolated two-sided assessment. After source-level deduplication, the assessor receives only the frozen bundle ā¬cB_c. It first assigns vtimācompatible,incompatible,unknownv_timā\compatible,incompatible,unknown\ from recorded times, granularities, and source-linked spans; coarse times remain unknown, and an incompatible order contributes to refutation only when occurrences, scope, timestamps, and sources are jointly resolved. Graph-native path order carries no evidential force of its own. It then scores admissible evidence for and against the specified relation on independent axes: S+ S^+ =Ī±timāStim++Ī±numāSnum++Ī±jdgāSjdg+, = _timS_tim^++ _numS_num^++ _jdgS_jdg^+, Sā S^- =Sjdgā, =S_jdg^-, (6) where Stim+S_tim^+ measures source-grounded temporal compatibility, Snum+S_num^+ numerical evidence, and SjdgĀ±S_jdg^Ā± assessor-identified supporting and refuting evidence. The refutation axis is conservative: incompatibility raises SāS^- only on explicit, source-grounded counter-evidence, each citing the item grounding it. The axes are evidence strengths, not complementary probabilities, so evidence on both sides raises both while unresolved evidence raises neither. The weights are development-split operating points, frozen at Ī±tim=0.45 _tim=0.45, Ī±num=0.22 _num=0.22, and Ī±jdg=0.33 _jdg=0.33, and are never re-calibrated per dataset or per protocol. Thresholds selected on the same split map the two axes to the verdict: y=supported,S+ā„Ļ+,Sā<Ļā,conflicting,S+ā„Ļ+,Sāā„Ļā,refuted,S+<Ļ+,Sāā„Ļā,insufficient,S+<Ļ+,Sā<Ļā.y= casessupported,&S^+ā„ _+,\;S^-< _-,\\ conflicting,&S^+ā„ _+,\;S^-ā„ _-,\\ refuted,&S^+< _+,\;S^-ā„ _-,\\ insufficient,&S^+< _+,\;S^-< _-. cases (7) For binary benchmarks the three non-supported verdicts are collapsed after four-way inference and solely for scoring, retaining verdict, axes, citations, and failure type for audit. With model revision, prompt, and decoding fixed, assessment satisfies fixed-bundle independence: ā¬c(1)=ā¬c(2)ā¹AssessĪøā(ā¬c(1))=AssessĪøā(ā¬c(2)).B_c^(1)=B_c^(2) Assess_Īø(B_c^(1))= Assess_Īø(B_c^(2)). (8) Compositional and hidden-intermediate queries. We extend pairwise verification to compositional queries Ļ=(e0,r1,e1,ā¦,rh,eh)Ļ=(e_0,r_1,e_1,ā¦,r_h,e_h) with hā„2hā„ 2, where each eie_i is a required occurrence and each rir_i the relation between eiā1e_i-1 and eie_i. Such a query is supported only if a scope-consistent chain grounds every occurrence and supplies admissible evidence for every adjacent relation; recovering the nodes or their graph order is insufficient, since each transition must be assessed from its bound textual, numerical, or temporal evidence. The hidden-intermediate setting replaces Ļ with qhid=(e0,eh,Ļ1:h,Ļ1:hā1)q_hid=(e_0,e_h, _1:h, _1:h-1), withholding the intermediate identities and leaving only endpoints, relation sequence, and type constraints. We therefore separate structural recovery, which identifies a candidate chain, from evidential verification, which additionally requires admissible evidence and citations for every relation. Retrieval-only procedural memory. As an optional extension, ā³M is a frozen bank of patient-independent retrieval policies derived from recurring training and development search failures; it holds no patient identifiers or facts, event values, evidence spans, provenance, labels, answers, or verdicts. Given a query and structured failure f, the controller may retrieve Ļ=RetrievePolicyā(ā³,c,f)Ļ= RetrievePolicy(M,c,f) to revise expansion, query selection, or replanning. Memory therefore changes only which admissible evidence is recovered, leaving Eq. (8) intact. Task Construction Existing clinical resources annotate relations or answer questions; none scores whether a record supplies admissible evidence for a specified relation. We therefore construct ten protocols from three annotated corpora and state their sources, negatives, and leakage controls here. Query sources and labels. Pairwise queries come from i2b2 2012 TLINK annotations (T1), n2c2 2018 medicationāADE annotations (T2), and recorded event order in MIMIC-IV structured tables (T3) (29; 12; 15; 16). In each case the annotation supplies only the query endpoints and the held-out label. Relation direction never enters events, edges, contexts, prompts, retrieval scores, or memory, so no method can read a verdict from the artifact it searches. A release audit verifies this before evaluation: the selectable edge set contains no support relation, gold direction appears only in evaluation labels, positive and negative candidates share one graph schema, and train/dev/test patients do not overlap. Negatives. A negative differing structurally from a positive can be solved by surface statistics rather than evidence, so each negative is a relation perturbation of a positive chain over the same events and contexts. Measured from the protocol alone, with no model involved, positives and negatives on T9 i2b2 have chain lengths 4.05 and 3.90, hops 3.05 and 2.90, and 8.21 and 8.74 contexts per queried endpoint; on T9 n2c2 the classes are identical in both. Neither carries a retrieval-side signature. The ten protocols. T1āT3 are the pairwise tasks defined above: given two named events, decide whether the record supports, refutes, conflicts on, or is insufficient for the stated relation. The other seven isolate one capability each. T4 scores retrieval alone, asking which node-attached contexts a method recovers before any verdict is formed. T5 masks the event, context, provenance, or timing that would establish a supported verdict, so the correct answer becomes insufficient or conflicting and a method still answering supported is answering from relevance. T6 leaves the evidence intact but adds four distractor strata: other-patient events, other-patient events sharing a name, other-patient contexts carrying the exact query terms, and same-patient contexts from a query-incompatible encounter. T7 runs the system with and without procedural memory on identical identifiers. T8 is a patient-disjoint latest-time split testing transfer rather than memorization. T9 states every queried event and asks whether all adjacent relations hold jointly, so one refuted transition refutes the query. T10 withholds the intermediate identities and scores whether a reconstructed chain is also verified and source-bound; it additionally runs on LUNGUAGE (21). What is held fixed. Every method receives the same frozen artifacts, identifiers, patient filters, encoders, candidate caps, context budget, and assessor; only the online retrieval operator differs. No protocol is filtered by any systemās output. Experiments Setup Baselines and implementation. B1 and B2 are direct and text-only retrieval; B3āB5 add the patient graph, a loose hybrid, and coupled graphācontext retrieval without isolated assessment; B6 is the full system and B7 adds procedural memory. We reproduce the online retrieval operators of MedGraphRAG, EHR-RAG, HippoRAG 2, KG2RAG, LightRAG, and GRAG on the shared index rather than importing published values. Qwen3-8B performs parsing and control, and the development-selected Qwen3-32B assesses frozen bundles (24); defaults are dmax=Rmax=5d_ =R_ =5 and 12 final contexts. (a) Primary results Discrimination and evidence recovery Evidence integrity Method T1 T2 T3 Ctx P@10ā”1/2/3^1/2/3_\; Pair R1/2/3 T5 T6 FSRā Sup.-rec B1 Direct LLM 46.1/33.0 62.4/62.4 50.2/40.4 ā ā 46.6/42.6 48.1/33.6 11.5 7.6 B2 Text-RAG 50.0/37.5 51.3/36.2 53.0/47.3 ā ā 21.0/29.4 48.3/39.2 96.6 93.1 B3 KG-only 47.8/44.3 48.7/48.5 66.4/63.2 ā ā 43.0/41.6 50.3/42.5 20.7 21.4 B4 Loose hybrid 51.7/49.9 50.4/49.7 62.6/62.3 ā ā 43.9/54.9 47.8/47.4 70.1 65.6 B5 Graphātext 50.1/35.9 39.4/33.9 50.2/35.1 ā ā 20.8/29.0 51.9/43.6 93.1 96.9 B6 Isolated 78.6/78.5 67.3/65.7 96.8/96.8 29.5/27.3/19.4 68.5/53.5/71.6 92.2/94.4 75.5/75.0 25.3 76.3 B7 Full 78.6/78.5 67.3/65.7 96.8/96.8ā 29.5/27.3/19.4 68.5/53.5/71.6 92.2/94.4 76.7/76.3 25.3 78.6 (b) Patient-cluster inference Protocol B6 BA [95% CI] B6 F1 [95% CI] Strongest B1āB5 Ī Holm pBAp_BA Holm pF1p_F1 T1 i2b2 78.6 [75.4, 81.8] 78.5 [75.2, 81.7] B4 (51.7) +26.9 .0074 .0074 T2 n2c2ā” 67.3 [62.0, 72.8] 65.7 [59.2, 72.2] B1 (62.4) +4.9 1.000 1.000 T3 MIMIC 96.8 [95.2, 98.2] 96.8 [95.2, 98.2] B3 (66.4) +30.4 .0074 .0074 T5 masked 92.2 [87.7, 96.5] 94.4 [91.0, 97.5] B1 (46.6) +45.6 .0074 .0074 T6 scope 75.5 [64.1, 87.4] 75.0 [63.3, 86.2] B5 (51.9) +23.6 .0130 .0074 T9 MIMIC 61.1 [57.6, 65.1] 59.8 [56.0, 64.1] B5 (51.9) +9.2 .0074 .0074 Table 1: Percent BA/F1 unless noted; superscripts 1/2/3 index T1/T2/T3, and FSR and supported recall are from T6. Bold marks the best value in a column, including where a control attains it: B1 has the lowest FSR and B5 the highest supported recall, each at the otherās expense. Panel (b) gives 95% bootstrap intervals and Holm-adjusted p-values. ā B7 equals B6 on T3 by construction; ā”exploratory. ā”Ctx P@10 has a fixed denominator of 10 while a query has only 5.11, 3.18, and 2.85 gold contexts on average, capping it at 49.1, 31.8, and 28.5; the reported values reach 60%, 86%, and 68% of that cap. Method T1/T2/T3 BA MedGraphRAG 72.3/54.4/78.8 EHR-RAG 56.5/56.6/70.0 HippoRAG 2 57.9/61.9/64.8 KG2RAG 61.0/56.6/70.2 LightRAG 71.3/52.2/76.8 GRAG 57.7/56.2/68.4 MedEventGraph-RAG 78.6/67.3/96.8 Table 2: Matched-index retrieval. Artifacts, identifiers, context budget, and assessor are held fixed and only the online retrieval operator changes; bold marks the best value per task. Inference cost is reported in the supplement. Metrics. Balanced accuracy (BA) and macro-F1 are primary for labeled protocols. The false support rate (FSR) is always reported together with supported recall, so that indiscriminate abstention cannot appear safe. T10 reports context recall (Ctx-R), hidden-bridge recall (Bridge), complete-chain recall at 20 (C@20), verified-chain recall (Verif), and provenance-valid chain recall (Prov.). Confidence intervals use 10,000 patient-cluster bootstrap replicates, and paired randomization tests use 10,000 permutations with Holm correction over the complete comparison family. T9 i2b2, T9 n2c2, and T10 n2c2 (n=11n=11) are exploratory because of their small patient-cluster counts. Stage BA F1 FSRā Ctx-R Chain-R E1 bound graphācontext 70.0 69.9 23.3 46.0 52.5 E2 + two-sided 72.5 72.2 16.7 63.1 47.5 E3 + isolated axes 75.0 74.7 15.0 63.1 47.5 E4 + replanning 75.8 75.7 15.8 60.0 55.0 E5 + memory 76.2 76.1 15.8 60.0 55.0 Table 3: Cumulative ablation, percent, on a frozen balanced 240-query set; Chain-R is complete-chain recall. Stages add search over the patient graph with its occurrence-bound context index (E1), source-bound two-sided retrieval (E2), evidence-isolated dual-axis verification (E3), verifier-guided replanning (E4, our B6), and retrieval-only procedural memory (E5, our B7), under the same candidate contract, assessor, and retrieval budget. E1āE3 decide from a coupled boundary and E4āE5 from an isolated one; E1āE3 hold the candidate superset fixed, while E4āE5 may extend it because that adaptive change is the treatment measured. (a) T9: all events named Method i2b2 n2c2 MIMIC B2 Text-RAG 52.5/42.0 50.0/33.3 50.2/49.5 B3 KG-only 47.5/32.2 50.7/50.7 50.3/34.9 B4 Loose hybrid 67.5/67.0 47.9/35.6 49.4/42.8 B5 Graphātext 50.0/33.3 50.0/33.3 51.9/51.9 Ours 82.5/82.2 66.4/65.3 61.1/59.8 (b) T10: intermediates withheld Method Ctx-R Bridge C@20 Verif Prov. B1 n/a n/a n/a n/a n/a B2 59.6 n/a n/a n/a n/a B3 n/a 56.1 5.8 n/a n/a B4 59.6 56.1 5.8 n/a n/a B5 50.2 90.1 67.1 n/a n/a B6 65.2 99.9 77.1 77.9 57.9 B7 65.2 99.9 77.1 77.9 58.8 Ours-L 92.3 98.8 77.1 80.7 70.0 Table 4: Compositional queries. (a) Percent BA/F1 on debiased balanced splits (n=40/140/620n=40/140/620). (b) Evidence survival from context and bridge recovery through complete candidate (C@20), verification, and provenance. Each metric in (b) is defined only for architectures that emit the corresponding output, and n/a marks a stage a method lacks: B1 retrieves nothing, B2 builds no graph, B3 attaches no contexts, and no B1āB5 variant emits a verification path or a source-bound chain. We write n/a rather than 0 because the quantity is undefined for them, not measured at zero, and claim nothing about an extended variant. Bold marks the best among B1āB7; Ours-L is LUNGUAGE, listed below the rule rather than compared in-column. Results Four tables support the analysis, each controlling a different confound. Table 1 reports the pairwise protocols T1āT3 together with the evidence-integrity protocols T5 and T6, placing accuracy beside a false-support column so that neither can be raised alone. Table 2 asks whether any gain is an artifact of our index rather than our search policy, so it holds artifacts, identifiers, context budget, and assessor fixed and replaces only the online retrieval operator with that of six published systems, reproduced on the shared index rather than imported from their papers. Table 3 attributes the gain internally by adding one evidence-layer mechanism at a time over a frozen balanced 240-query set, so that a change in what is retrieved can be separated from a change in how it is judged. Table 4 moves from pairwise to compositional queries, with every event named in panel (a) and the intermediate identities withheld in panel (b). Does broad, source-linked discovery recover evidence that flat and graph-only retrieval miss? B6 gains 26.9, 4.9, and 30.4 BA over the strongest matched control (Table 1). Patient-cluster inference confirms T1 and T3 after Holm correction (p=.0074p=.0074); T2 is exploratory, and T3 measures recorded order rather than causality. Under the matched index the best external result per task is 72.3, 61.9, and 78.8, attained by three different systems, against our 78.6, 67.3, and 96.8 (Table 2), so what differs is the search policy rather than the index. Two measurements bound the result: T4 reaches 89.5 context recall, so the occurrence-linked contexts are retrievable, and T8 retains 71.5/69.5 on a patient-disjoint split, so the gain survives distribution shift. Does relevance establish evidential direction? It does not, and balanced accuracy alone conceals this: under T5 a system that answers from relevance must produce false support. The retrieval-coupled controls B2 and B5 do exactly that, reporting false support on 96.6% and 93.1% of non-supported queries: they answer supported almost whenever retrieval returns anything. The opposite failure is equally available, since B1 holds FSR to 11.5 only by abstaining almost everywhere, at 7.6% supported recall. B6 is the only configuration that separates the two axes, reaching 92.2/94.4 BA/F1 with zero false support on T5 (+45.6+45.6 BA over B1, p=.0074p=.0074) while holding FSR at 25.3 with 76.3% supported recall overall. One case shows the mechanism: asked whether kidney transplantation preceded pancreas transplantation once the establishing context was masked, B5 returned supported from residual graph linkage, while B6 returned insufficient and named the missing provenance. The graph proposes the pair; it cannot substitute for the source that orders it. Does admissibility buy safety by discarding evidence? It does not. Under T6 B6 reaches 75.5/75.0 (+23.6+23.6 BA over B5, p=.0130p=.0130) while keeping 76.3% supported recall, so the patient boundary holds without rejecting legitimate cross-encounter evidence; blanket rejection would have depressed both quantities. The cumulative ablation (Table 3) separates the two routes to that result. E2 acts on what is found: context recall rises from 46.0 to 63.1 and FSR falls from 23.3 to 16.7. E3 then acts only on how that evidence is judged, leaving context and chain recall unchanged at 63.1 and 47.5 while FSR falls to 15.0 and BA rises to 75.0. Withholding an unfounded decision and failing to retrieve evidence are thus separable, and here the boundary costs no accuracy. Verifier-guided replanning (E4) then lifts chain recall to 55.0 at 75.8 BA, so the boundary does not foreclose further search. Retrieval-only procedural memory (E5) adds 0.4 BA/F1 at unchanged FSR (15.8), context recall (60.0), and chain recall (55.0). Where does the compositional pipeline still fail? T9 exceeds the strongest B2 to B5 control on every corpus (Table 4a), though only the larger MIMIC split supports a corrected result (+9.2+9.2 BA, p=.0074p=.0074). T10 explains why (Table 4b). Read as a funnel from bridge recovery through complete candidate at 20, verification, and provenance, the comparison is one of capability before degree. B5 reaches 90.1 and 67.1, so structural reach is attainable with no grounding stage at all. The substantive result is that our system populates the last two columns at all, and non-trivially: B6 lifts the funnel to 99.9 and 77.1, preserves 77.9 through verification, yet retains only 57.9 as provenance-valid. Structural reach is therefore solved while source binding is not, and the 20.0-point drop from verification to provenance localizes the bottleneck. Ours-L shows the bottleneck is contextual rather than architectural: richer contexts raise Ctx-R from 65.2 to 92.3 and provenance to 70.0 at unchanged C@20. Can procedural memory improve retrieval without reaching judgment? B7 stores no patient facts, spans, answers, or verdicts. It improves T6 by 1.2/1.3 BA/F1 through three corrected retrieval failures at unchanged FSR (25.3), so the gain enters through what is found rather than how it is judged, and equals B6 exactly on T1āT3 and T5. Fixed-bundle replay with memory disabled, enabled, or injected with fact-like content changed no verdict and no (S+,Sā)(S^+,S^-). What does the boundary cost? Evidence isolation is materially more expensive: B6 uses 4.625 calls and 12.8k tokens per query against 1.000 and 1.4k for B2, a 4.6Ć4.6Ć call and 9.3Ć9.3Ć token premium, while procedural memory adds 3.06 seconds of local work and no LLM call. Per-method figures are in the supplement. Discussion and Conclusion Longitudinal verification fails in three independent ways, each answered by a design decision rather than by scale. Binding the unit of search to an event occurrence and its source steers discovery by what still needs grounding, lifting hidden-bridge recovery to 99.9; scoring support and refutation on independent axes yields zero false support under masked evidence, against 96.6 and 93.1 for the coupled controls; and a contract applied at assessment time holds the patient boundary under distractors while retaining 76.3% supported recall. These answers are separable, and the residual gap is localized: we verify 77.9 of reconstructed chains but ground 57.9. The principle generalizes beyond medicine: a verdict-producing system should separate the signals deciding where to look from the evidence permitted to decide, though we claim neither causality nor clinical safety beyond the corpora tested. References Asai et al. (2024) A. Asai, Z. Wu, Y. Wang, A. Sil, and H. Hajishirzi Self-RAG: learning to retrieve, generate, and critique through self-reflection. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR), Cited by: Related Work. Bethard et al. (2016) S. Bethard, G. Savova, W. Chen, L. Derczynski, J. Pustejovsky, and M. Verhagen SemEval-2016 task 12: clinical TempEval. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016), External Links: Document Cited by: Introduction. Cao et al. (2026) L. Cao, Q. Chen, and Y. Guo EHR-RAG: bridging long-horizon structured electronic health records and large language models via enhanced retrieval-augmented generation. arXiv preprint arXiv:2601.21340. Cited by: Related Work. Chapman et al. (2001) W. W. Chapman, W. Bridewell, P. Hanbury, G. F. Cooper, and B. G. Buchanan A simple algorithm for identifying negated findings and diseases in discharge summaries. Journal of Biomedical Informatics 34 (5), p. 301ā310. External Links: Document Cited by: Introduction. Chen et al. (2023) Z. Chen, A. H. Cano, A. Romanou, A. Bonnet, K. Matoba, F. Salvi, M. Pagliardini, S. Fan, A. Kƶpf, A. Mohtashami, A. Sallinen, A. Sakhaeirad, V. Swamy, I. Krawczuk, D. Bayazit, A. Marmet, S. Montariol, M. Hartley, M. Jaggi, and A. Bosselut MEDITRON-70B: scaling medical pretraining for large language models. arXiv preprint arXiv:2311.16079. Cited by: Related Work. Edge et al. (2024) D. Edge, H. Trinh, N. Cheng, J. Bradley, A. Chao, A. Mody, S. Truitt, and J. Larson From local to global: a graph RAG approach to query-focused summarization. arXiv preprint arXiv:2404.16130. Cited by: Related Work. Guo et al. (2024) Z. Guo, L. Xia, Y. Yu, T. Ao, and C. Huang LightRAG: simple and fast retrieval-augmented generation. arXiv preprint arXiv:2410.05779. Cited by: Related Work. GutiĆ©rrez et al. (2024) B. J. GutiĆ©rrez, Y. Shu, Y. Gu, M. Yasunaga, and Y. Su HippoRAG: neurobiologically inspired long-term memory for large language models. In Advances in Neural Information Processing Systems 37 (NeurIPS), Cited by: Related Work. GutiĆ©rrez et al. (2025) B. J. GutiĆ©rrez, Y. Shu, W. Qi, S. Zhou, and Y. Su From RAG to memory: non-parametric continual learning for large language models. In Proceedings of the 42nd International Conference on Machine Learning (ICML), Cited by: Related Work. Harkema et al. (2009) H. Harkema, J. N. Dowling, T. Thornblade, and W. W. Chapman ConText: an algorithm for determining negation, experiencer, and temporal status from clinical reports. Journal of Biomedical Informatics 42 (5), p. 839ā851. External Links: Document Cited by: Introduction. He et al. (2024) X. He, Y. Tian, Y. Sun, N. V. Chawla, T. Laurent, Y. LeCun, X. Bresson, and B. Hooi G-Retriever: retrieval-augmented generation for textual graph understanding and question answering. In Advances in Neural Information Processing Systems 37 (NeurIPS), Cited by: Related Work. Henry et al. (2020) S. Henry, K. Buchan, M. Filannino, A. Stubbs, and O. Uzuner 2018 n2c2 shared task on adverse drug events and medication extraction in electronic health records. Journal of the American Medical Informatics Association 27 (1), p. 3ā12. External Links: Document Cited by: Query sources and labels.. Hu et al. (2025) Y. Hu, Z. Lei, Z. Zhang, B. Pan, C. Ling, and L. Zhao GRAG: graph retrieval-augmented generation. In Findings of the Association for Computational Linguistics: NAACL 2025, p. 4145ā4157. External Links: Document Cited by: Related Work. Jiang et al. (2025) P. Jiang, C. Xiao, M. Jiang, P. Bhatia, T. Kass-Hout, J. Sun, and J. Han Reasoning-enhanced healthcare predictions with knowledge graph community retrieval. In Proceedings of the Thirteenth International Conference on Learning Representations (ICLR), Cited by: Related Work. Johnson et al. (2023a) A. E. W. Johnson, L. Bulgarelli, L. Shen, A. Gayles, A. Shammout, S. Horng, T. J. Pollard, S. Hao, B. Moody, B. Gow, L. H. Lehman, L. A. Celi, and R. G. Mark MIMIC-IV, a freely accessible electronic health record dataset. Scientific Data 10 (1), p. 1. External Links: Document Cited by: Query sources and labels.. Johnson et al. (2023b) A. E. W. Johnson, T. J. Pollard, S. Horng, L. A. Celi, and R. G. Mark MIMIC-IV-Note: deidentified free-text clinical notes (version 2.2). Note: PhysioNet External Links: Document Cited by: Query sources and labels.. Labrak et al. (2024) Y. Labrak, A. Bazoge, E. Morin, P. Gourraud, M. Rouvier, and R. Dufour BioMistral: a collection of open-source pretrained large language models for medical domains. In Findings of the Association for Computational Linguistics: ACL 2024, External Links: Document Cited by: Related Work. Li et al. (2024) S. Li, Y. He, H. Guo, X. Bu, G. Bai, J. Liu, J. Liu, X. Qu, Y. Li, W. Ouyang, W. Su, and B. Zheng GraphReader: building graph-based agent to enhance long-context abilities of large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, Cited by: Related Work. Liu et al. (2025) H. Liu, S. Wang, and J. Li Knowledge graph retrieval-augmented generation via GNN-guided prompting. In Proceedings of the Second Conference on Language Modeling (COLM), Cited by: Related Work. Ma et al. (2025) S. Ma, C. Xu, X. Jiang, M. Li, H. Qu, C. Yang, J. Mao, and J. Guo Think-on-graph 2.0: deep and faithful large language model reasoning with knowledge-guided retrieval augmented generation. In Proceedings of the Thirteenth International Conference on Learning Representations (ICLR), Cited by: Related Work. Moon et al. (2025) J. H. Moon, G. Choi, P. Rabaey, M. G. Kim, H. G. Hong, J. O. Lee, H. Yoon, E. Doe, J. Kim, H. Sharma, D. C. de Castro, J. A. Valle, and E. Choi LUNGUAGE: a benchmark for structured and sequential chest x-ray interpretation. arXiv preprint arXiv:2505.21190. Cited by: The ten protocols.. Packer et al. (2023) C. Packer, S. Wooders, K. Lin, V. Fang, S. G. Patil, I. Stoica, and J. E. Gonzalez MemGPT: towards LLMs as operating systems. arXiv preprint arXiv:2310.08560. Cited by: Related Work. Park et al. (2026) H. Park, J. Seo, J. Mun, H. Park, W. Byeon, S. J. Kim, H. Im, J. Lee, and S. Kim M3^3KG-RAG: multi-hop multimodal knowledge graph-enhanced retrieval-augmented generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: Related Work. Qwen Team (2025) Qwen Team Qwen3 technical report. arXiv preprint arXiv:2505.09388. Cited by: Baselines and implementation.. Sarthi et al. (2024) P. Sarthi, S. Abdullah, A. Tuli, S. Khanna, A. Goldie, and C. D. Manning RAPTOR: recursive abstractive processing for tree-organized retrieval. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR), Cited by: Related Work. Shinn et al. (2023) N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao Reflexion: language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems 36 (NeurIPS), Cited by: Related Work. Shurrab et al. (2026) S. Shurrab, M. Al-Omari, D. E. Samad, and F. E. Shamout EHR-RAGp: retrieval-augmented prototype-guided foundation model for electronic health records. arXiv preprint arXiv:2605.12335. Cited by: Related Work. Styler IV et al. (2014) W. F. Styler IV, S. Bethard, S. Finan, M. Palmer, S. Pradhan, P. C. de Groen, B. Erickson, T. Miller, C. Lin, G. Savova, and J. Pustejovsky Temporal annotation in the clinical domain. Transactions of the Association for Computational Linguistics 2, p. 143ā154. External Links: Document Cited by: Introduction. Sun et al. (2013) W. Sun, A. Rumshisky, and O. Uzuner Evaluating temporal relations in clinical text: 2012 i2b2 challenge. Journal of the American Medical Informatics Association 20 (5), p. 806ā813. External Links: Document Cited by: Query sources and labels.. Uzuner et al. (2011) O. Uzuner, B. R. South, S. Shen, and S. L. DuVall 2010 i2b2/VA challenge on concepts, assertions, and relations in clinical text. Journal of the American Medical Informatics Association 18 (5), p. 552ā556. External Links: Document Cited by: Introduction. Wang et al. (2024) Z. Z. Wang, J. Mao, D. Fried, and G. Neubig Agent workflow memory. arXiv preprint arXiv:2409.07429. Cited by: Related Work. Wrenn et al. (2010) J. O. Wrenn, D. M. Stein, S. Bakken, and P. D. Stetson Quantifying clinical narrative redundancy in an electronic health record. Journal of the American Medical Informatics Association 17 (1), p. 49ā53. External Links: Document Cited by: Introduction. Wu et al. (2025) J. Wu, J. Zhu, Y. Qi, J. Chen, M. Xu, F. Menolascina, Y. Jin, and V. Grau Medical graph RAG: evidence-based medical large language model via graph retrieval-augmented generation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 28443ā28467. External Links: Document Cited by: Related Work. Yan et al. (2024) S. Yan, J. Gu, Y. Zhu, and Z. Ling Corrective retrieval augmented generation. arXiv preprint arXiv:2401.15884. Cited by: Related Work. Zhao et al. (2024) A. Zhao, D. Huang, Q. Xu, M. Lin, Y. Liu, and G. Huang ExpeL: LLM agents are experiential learners. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38, p. 19632ā19642. External Links: Document Cited by: Related Work. Zhao et al. (2025) X. Zhao, S. Liu, S. Yang, and C. Miao MedRAG: enhancing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot. In Proceedings of the ACM Web Conference 2025 (W), p. 4442ā4457. External Links: Document Cited by: Related Work. Zhu et al. (2025) X. Zhu, Y. Xie, Y. Liu, Y. Li, and W. Hu Knowledge graph-guided retrieval augmented generation. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), p. 8912ā8924. External Links: Document Cited by: Related Work.