Consistent Relexicalization of Clinical Documents using Graph-Based Approach
Dipankar Das, Atri Mandal, Sandeep Singh, Tushar Shandhilya
cs.CL · cs.AI
Quick verdict
G-RELIC pairs GPT-4.1 with a clinical knowledge graph for consistent de-ID relexicalization; large gains, but small proprietary corpus, no statistical tests.
Summary
The paper introduces G-RELIC, a relexicalization framework from Oracle Health AI that replaces PHI/PII in clinical documents with consistent surrogates. It targets two failure modes of independent, per-document replacement: identity fragmentation/merging across longitudinal records, and temporal incoherence (ordering, intervals, weekday/calendar consistency, medical plausibility). The pipeline has three stages: NER plus LLM entity identification and role classification; construction of a Clinical Knowledge Graph (Neo4j) anchored to patient IDs; and multi-tier entity resolution combining deterministic primary-identifier matching with weighted secondary-attribute similarity and thresholded linkage. Temporal entities are shifted by a patient-specific offset validated by a rule layer (ICD-10-derived constraints) plus an LLM confidence check with fallback recalibration. On 400 documents from 50 patients (Gold/Silver+/Silver tiers, mostly proprietary production data), G-RELIC reports 92.5% entity consistency versus 62.1% for a prompt-based baseline (+30.4pp) and 91.9% temporal consistency versus 46.0% (+45.9pp), while keeping de-identification precision/AoN recall roughly on par with RedactOR on proprietary and N2C2 2014 data, with 83-84% token reductions. Limitations: no statistical significance tests or confidence intervals; a trivial persistent-mapping baseline is absent; the 'random shift' baseline appears weaker than standard per-patient constant offsets; the dataset is small and largely proprietary; no code release; and the LLM validation layer sits awkwardly with the 'deterministic' claim. A 7-day production deployment (1.18M requests) is also reported.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A graph-based one-to-one entity mapping over a Clinical Knowledge Graph raises entity consistency to 92.5% versus 62.1% for prompt-based relexicalization (+30.4 percentage points).
- Patient-specific deterministic temporal offsets, validated by rule and LLM layers, raise temporal consistency to 91.9% versus 46.0% for the prompt baseline (+45.9 percentage points).
- G-RELIC matches or slightly exceeds the RedactOR-style prompt baseline's de-identification precision/AoN recall on both a proprietary corpus and N2C2 2014.
- A single-pass GPT-4.1 design cuts input tokens by 83%, output tokens by 84%, and P90 latency by 65% versus the multi-pass RedactOR strategy.
- The system ran in production for 7 days, processing 1,176,642 relexicalization requests at a 98.7% success rate for $105 of LLM spend.
Strengths
Clear failure taxonomy and motivating example
Section I-A distinguishes identity fragmentation vs identity merging and four temporal error classes; Table I's before/after example (patient 'A Miller' splitting into 'J Clarke' and 'Michael Johnson'; a 43-year-old still described as at his '50th birthday') makes the consistency problem concrete and testable.
Fully specified entity resolution procedure
Section III-C spells out blocking on first/last name with a Kronecker-delta gender gate, Jaro-Winkler-type similarity, weighted attribute scoring R(u,v), threshold γ, ontological validation, and logged manual review of failures — implementation-level detail rarely disclosed in industry papers.
Temporal consistency decomposed into checkable properties
Honest confound attribution on privacy
Production-scale operational telemetry
Weaknesses
No statistical rigor
Missing the obvious baseline
Possible strawman temporal baseline
Determinism overstated; privacy-utility tension unmeasured
Closed data and code; a stated target is missed
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
G-RELIC (Oracle Health AI; arXiv 2609.21387) is a method paper on relexicalization — replacing PHI/PII in clinical documents with realistic surrogates — aimed at cross-document consistency in longitudinal records. The abstract stakes three quantitative claims: entity consistency 92.5% vs 62.1% for a prompt-based baseline (+30.4pp), temporal consistency 91.9% vs 46.0% (+45.9pp), and maintained privacy benchmarks, plus 83%/84% token and 65% P90 latency reductions versus a multi-pass competitor. Section I-A builds a two-part failure taxonomy (identity fragmentation vs merging; four temporal inconsistency classes) and the framework is billed as mathematically bounded and deterministic.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is relexicalization in clinical NLP?
Relexicalization replaces protected health information (PHI/PII) in clinical text with realistic surrogate values — fake names, shifted dates — instead of deleting or masking it, so de-identified datasets keep clinical usefulness for research while protecting patient privacy.
What is G-RELIC and how does it improve consistency?
G-RELIC (Graph-Based Contextual Relexicalization with Improved Consistency) is an Oracle Health AI framework that combines LLM-based entity detection with a clinical knowledge graph, so the same real-world entity is always replaced by the same surrogate across a patient's records, and dates are shifted with patient-specific offsets that preserve ordering, intervals, weekdays, and medical plausibility.
How much better is G-RELIC than prompt-based de-identification?
On the paper's 400-document corpus (50 patients, mostly proprietary data), it reports 92.5% entity consistency versus 62.1% for a prompt-based baseline (+30.4 points) and 91.9% versus 46.0% temporal consistency (+45.9 points), with de-identification precision/recall roughly matching RedactOR on proprietary and N2C2 2014 data.
Is the G-RELIC code or dataset publicly available?
No code release is announced. The main corpus includes proprietary production EHR data from 12 provider sites; only the public N2C2 2014 dataset was used for the privacy comparison, and the published temporal-validation prompt is the main reusable artifact.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–3 (architecture, CKG structure, entity-resolution decision logic) are referenced but not analyzable from the extracted text.
- All production telemetry in Section V-E is self-reported and independently unverifiable.
- The 'Rule-Based [11]' baseline's implementation cannot be verified from its citation (an entity-resolution matching paper).
- Results are pooled across the three dataset tiers; per-tier breakdowns are absent (footnote 1).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.