跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2608.03172 · v6

Surrogate Substitution Preserves PHI Detectability: A Multi-Detector Equivalence Study

首次提交 2026-08-04 14:04
版本更新 2026-11-18 17:58
雷达观测 2026-11-19 15:04
arXiv ↗PDF ↗

摘要

Structure-preserving de-identification replaces protected health information (PHI) with realistic same-type surrogates ("Anna S." becomes "Maria S.", not [NAME]). De-identified notes are what allow clinical text to be shared for model development, and many pipelines run a second PHI detector over them as a safety net. We ask a narrow, testable question: on the spans a de-identifier actually replaces, can downstream PHI detectors still find the surrogate as well as they found the original value? We call this detectability preservation and treat it as a proxy, not as general utility. We propose a paired protocol that (i) scores the same spans before and after replacement with an identical panel of 11 detectors from three families, (ii) measures the substitution effect only on replaced spans, keeping it separate from coverage, and (iii) bounds the effect with document-level cluster-bootstrap equivalence testing and posterior probabilities instead of significance tests. Applied to a commercial transform on seven benchmarks in seven languages (1,750 documents), pooled recall on replaced spans moves from 75.9% to 74.7% (Δ=−1.24Δ=-1.24 points, 90% CI [−1.65,−0.84][-1.65,-0.84]), within a ±2\pm2-point margin under overlap matching. The effect is not uniform: it exceeds ±2\pm2 points for several detectors, for two identifier-dense non-English benchmarks, and under exact-boundary matching. Two-rater annotation of surrogates (κ=0.79κ=0.79), a same-slot counterfactual, and detector test-retest show that the residual has both generator-side and detector-side causes. On ICD-10 diagnosis classification of the same clinical documents, accuracy changes by at most 1 point under substitution, Faker surrogates, or redaction.

作者备注:Accepted to the Machine Learning for Health (ML4H) Symposium 2026, Proceedings track (PMLR 346), held in Sydney alongside NeurIPS 2026. 14 pages, 3 figures, 12 tables. Code, data and per-span predictions: https://github.com/Custodian-Labs/guardian-layer-phi-benchmark ; dashboard: https://custodianlabs.pages.dev

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v6,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v6 · 2026-11-18 17:58 · 最新观测版本

Surrogate Substitution Preserves PHI Detectability: A Multi-Detector Equivalence Study

Structure-preserving de-identification replaces protected health information (PHI) with realistic same-type surrogates ("Anna S." becomes "Maria S.", not [NAME]). De-identified notes are what allow clinical text to be shared for model development, and many pipelines run a second PHI detector over them as a safety net. We ask a narrow, testable question: on the spans a de-identifier actually replaces, can downstream PHI detectors still find the surrogate as well as they found the original value? We call this detectability preservation and treat it as a proxy, not as general utility. We propose a paired protocol that (i) scores the same spans before and after replacement with an identical panel of 11 detectors from three families, (ii) measures the substitution effect only on replaced spans, keeping it separate from coverage, and (iii) bounds the effect with document-level cluster-bootstrap equivalence testing and posterior probabilities instead of significance tests. Applied to a commercial transform on seven benchmarks in seven languages (1,750 documents), pooled recall on replaced spans moves from 75.9% to 74.7% (Δ=−1.24Δ=-1.24 points, 90% CI [−1.65,−0.84][-1.65,-0.84]), within a ±2\pm2-point margin under overlap matching. The effect is not uniform: it exceeds ±2\pm2 points for several detectors, for two identifier-dense non-English benchmarks, and under exact-boundary matching. Two-rater annotation of surrogates (κ=0.79κ=0.79), a same-slot counterfactual, and detector test-retest show that the residual has both generator-side and detector-side causes. On ICD-10 diagnosis classification of the same clinical documents, accuracy changes by at most 1 point under substitution, Faker surrogates, or redaction.

查看此版本解读 →查看此版本原文 ↗