From 22c23272bbe03961f7d717e0de42955a073729de Mon Sep 17 00:00:00 2001 From: gglee Date: Mon, 31 Aug 2026 16:17:03 +0900 Subject: [PATCH] =?UTF-8?q?fix(BIOP02-144):=20HPV=20site-label=20Cram?= =?UTF-8?q?=C3=A9r's=20V=200.378=20=E2=86=92=200.397=20(14=EA=B3=B3)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 정본: experiments/crosscancer/site_audit/BIOP02-123_FINDINGS.md:26 HPV 0.397 · BRAF 0.333 · EGFR 0.36 · KRAS 0.36 · MSI 0.288 0.378은 원고에서 두 가지 서로 다른 수치로 쓰인다. 일괄치환하면 원고가 깨지므로 줄 단위로 지정해 정정했다. 정정(HPV site-label V) 14곳: CONF_ABSTRACT_GIW2026_BIOP02.md:57 DRAFT_paperC_sentence-skeleton_ko.md:74,151 LONG_ABSTRACT_GIW2026_BIOP02_{ko,en}.md:44 DRAFT_paperC_full_{ko,en}.md:64,147 DRAFT_ML4H_v2_full_ko.md:72,169 · DRAFT_ML4H_v2_full.md:70,167 DRAFT_ML4H_v2_CHANGELOG.md:37 보존(내분비요법 라우팅 비용, 무관한 수치) 6곳: sections/02_results.md:46 · DRAFT_paperC_sentence-skeleton_ko.md:98 DRAFT_paperC_full_{ko,en}.md:90 · DRAFT_ML4H_v2_full{_ko,}.md:100,98 출처: experiments/kkkim/EXPERIMENTS_INDEX_2026-07.md:27 (endocrine 0.378) 제출된 GIW 200단어 초록 본문에는 이 수치가 없어 제출본은 영향 없다. --- manuscript/CONF_ABSTRACT_GIW2026_BIOP02.md | 2 +- manuscript/DRAFT_ML4H_v2_CHANGELOG.md | 2 +- manuscript/DRAFT_ML4H_v2_full.md | 4 ++-- manuscript/DRAFT_ML4H_v2_full_ko.md | 4 ++-- manuscript/DRAFT_paperC_full_en.md | 4 ++-- manuscript/DRAFT_paperC_full_ko.md | 4 ++-- manuscript/DRAFT_paperC_sentence-skeleton_ko.md | 4 ++-- manuscript/LONG_ABSTRACT_GIW2026_BIOP02_en.md | 2 +- manuscript/LONG_ABSTRACT_GIW2026_BIOP02_ko.md | 2 +- 9 files changed, 14 insertions(+), 14 deletions(-) diff --git a/manuscript/CONF_ABSTRACT_GIW2026_BIOP02.md b/manuscript/CONF_ABSTRACT_GIW2026_BIOP02.md index 0fc4053..605d40f 100644 --- a/manuscript/CONF_ABSTRACT_GIW2026_BIOP02.md +++ b/manuscript/CONF_ABSTRACT_GIW2026_BIOP02.md @@ -54,7 +54,7 @@ AI 모델은 H&E 조직형태에서 분자 상태를 읽어낸다. 그러나 읽 1. **폐 KRAS-G12C 0.681 < 조직형만 쓰는 기준선 0.793** — 양성 14명이라 우리 규칙상 미결 축이므로, 결론이 아니라 관찰로만 쓴다. 2. **Yale pCR 0.533 [0.411, 0.653]** — 탐색적 결과 확인이며 독립 검증이 아니다. Farahmand 0.80에 못 미친다고 명시. -3. **Cramér's V = 0.378** (HPV site-라벨 구조) — 초록에서는 "3종 중 2종"만 남겼다. +3. **Cramér's V = 0.397** (HPV site-라벨 구조) — 초록에서는 "3종 중 2종"만 남겼다. 4. **음성 결과의 모델 무관 재현** (Lauren 실패·ERBB2 무신호가 3종 모두에서) — 원고 R5 근거. 초록에 넣기엔 자리가 없었다. 5. **operating point 미완성** — 오배정률 1.00의 해석이 임계에 의존하며 사전정의 분석(BIOP02-138)이 미완이라는 사실. 초록에서는 수치 해석을 안 하는 방식으로 우회했다. 6. 그림 우선순위: **Fig2(결정지도) + Fig4(검정력 천장)**. 지도보다 지도를 만든 증거가 먼저다. diff --git a/manuscript/DRAFT_ML4H_v2_CHANGELOG.md b/manuscript/DRAFT_ML4H_v2_CHANGELOG.md index 1983c25..a066f85 100644 --- a/manuscript/DRAFT_ML4H_v2_CHANGELOG.md +++ b/manuscript/DRAFT_ML4H_v2_CHANGELOG.md @@ -34,7 +34,7 @@ 8. **20-seed HPV/Virchow2 non-adoption stated (Discussion).** New: 20-seed indicated the Virchow2 HPV threshold stabilises (≈0.837) and *would* flip that cell to pass, but **not adopted** pending braveji (BIOP02-123); reported status stays "2 of 3 models". Source: `04_discussion.md` item 6; `experiments/kkkim/20260820_shuffle_null_20seed/`. -9. **Site-confounding + stain-scope co-located (Discussion).** Per review guidance, the site-audit limits (HPV V = 0.378; lung V = 1.000) and the stain-normalisation scope limit are placed in **one** Limitations paragraph, so it is clear that neither fully covers the headline HPV/lung results. +9. **Site-confounding + stain-scope co-located (Discussion).** Per review guidance, the site-audit limits (HPV V = 0.397; lung V = 1.000) and the stain-normalisation scope limit are placed in **one** Limitations paragraph, so it is clear that neither fully covers the headline HPV/lung results. 10. **Two anchor limitation blocks preserved** from `04_discussion.md` (BIOP02-75 critic gate): endpoint-specific baseline value, probability-level fidelity, HER2 honest negative, cell-line transfer. diff --git a/manuscript/DRAFT_ML4H_v2_full.md b/manuscript/DRAFT_ML4H_v2_full.md index cce8149..83b01c1 100644 --- a/manuscript/DRAFT_ML4H_v2_full.md +++ b/manuscript/DRAFT_ML4H_v2_full.md @@ -67,7 +67,7 @@ Every endpoint is reported alongside a shuffle-null (5-seed), a prevalence basel -† HPV robustness passed 5-seed chance-exclusion in only 2 of 3 foundation models (UNI and UNI2-h) and did not clear the pre-specified criterion in Virchow2 (real 0.9199 < threshold 0.9234, margin −0.0035 — a wide shuffle-null spread, not signal absence; see R5). The site audit also found site-label structuring (Cramér's V = 0.378). HPV is the single powered anchor that fixes one end of the map — not a generalisation of the law and not a model-independent confirmation. +† HPV robustness passed 5-seed chance-exclusion in only 2 of 3 foundation models (UNI and UNI2-h) and did not clear the pre-specified criterion in Virchow2 (real 0.9199 < threshold 0.9234, margin −0.0035 — a wide shuffle-null spread, not signal absence; see R5). The site audit also found site-label structuring (Cramér's V = 0.397). HPV is the single powered anchor that fixes one end of the map — not a generalisation of the law and not a model-independent confirmation. ‡ For lung histology (positive control) the site audit gave V(site, label) = 1.000: TCGA-LUAD/LUSC institution codes coincide 100% with histology, so the morphological signal cannot be separated from the site signature. This limit is stated explicitly wherever the positive control is interpreted. ### R2. Our pre-registered split cannot decide the mutation axes — the power ceiling @@ -164,7 +164,7 @@ In breast HER2, routing from predicted subtype failed consistently in this cohor The limits are placed in front. All results are retrospective, cohort-level and `hypothesis_only`; they are not claims of individual-level benefit. -**Site/batch confounding and stain-variation limits — the two "is it really morphology?" attacks, together.** The site-disjoint split prevented leakage in which slides from the same institution enter both training and evaluation, but the coupling between label and institution (confounding) remains, which limits interpreting performance as a pure morphological signal. A site/batch confounding audit confirmed in five endpoints that the site-disjoint split confounds labels with the tissue source site — lung histology (positive control) at V = 1.000, where morphology and site signature cannot be separated; head and neck HPV at V = 0.378; gastric Lauren, whose prevalence shifts from 0.46 to 0.88; and lung EGFR and KRAS, which also show significant site-label association. This is a necessary condition for confounding, not proof that "the model reads site". Separately, H&E stain normalisation was not applied in the main pipeline; a stain-normalisation robustness check preserved the breast anchor pattern (HER2 near-chance, ER high, PAM50 preserved), but **this check covers the breast anchor only** — the cross-cancer headline axes (head and neck HPV, lung histology) could not be re-checked because their raw slides were lost and re-extraction is deferred. The two headline results a reviewer would most likely attack as scanner/stain artefacts are therefore exactly the ones that carry the site-confounding flags and are not yet stain-verified; we state this openly. Site predictability from H&E and leave-one-site-out performance will adjudicate the confounding question; until then these limits are stated wherever the two anchors are interpreted. +**Site/batch confounding and stain-variation limits — the two "is it really morphology?" attacks, together.** The site-disjoint split prevented leakage in which slides from the same institution enter both training and evaluation, but the coupling between label and institution (confounding) remains, which limits interpreting performance as a pure morphological signal. A site/batch confounding audit confirmed in five endpoints that the site-disjoint split confounds labels with the tissue source site — lung histology (positive control) at V = 1.000, where morphology and site signature cannot be separated; head and neck HPV at V = 0.397; gastric Lauren, whose prevalence shifts from 0.46 to 0.88; and lung EGFR and KRAS, which also show significant site-label association. This is a necessary condition for confounding, not proof that "the model reads site". Separately, H&E stain normalisation was not applied in the main pipeline; a stain-normalisation robustness check preserved the breast anchor pattern (HER2 near-chance, ER high, PAM50 preserved), but **this check covers the breast anchor only** — the cross-cancer headline axes (head and neck HPV, lung histology) could not be re-checked because their raw slides were lost and re-extraction is deferred. The two headline results a reviewer would most likely attack as scanner/stain artefacts are therefore exactly the ones that carry the site-confounding flags and are not yet stain-verified; we state this openly. Site predictability from H&E and leave-one-site-out performance will adjudicate the confounding question; until then these limits are stated wherever the two anchors are interpreted. On model independence, the relative AUROC ordering among lung endpoints (histology > EGFR > KRAS) and the principal negative results were preserved across all three foundation models evaluated, whereas whether an individual molecular axis clears chance-exclusion varied by model. This is ordering stability, not model independence of the law as a whole: the headline single-endpoint confirmations (HPV, colorectal BRAF) each cleared chance-exclusion in only two of three foundation models. We note that a 20-seed re-check indicated the Virchow2 HPV threshold stabilises (to ≈0.837) and would flip that cell to a pass, but **this verdict change is not adopted** pending braveji's Critic confirmation (BIOP02-123); the reported status remains "2 of 3 models". diff --git a/manuscript/DRAFT_ML4H_v2_full_ko.md b/manuscript/DRAFT_ML4H_v2_full_ko.md index d992a29..a5ec5ed 100644 --- a/manuscript/DRAFT_ML4H_v2_full_ko.md +++ b/manuscript/DRAFT_ML4H_v2_full_ko.md @@ -69,7 +69,7 @@ H&E 조직 이미지에서 종양의 분자 상태를 예측하는 연구는 이 -† HPV 견고성은 3개 파운데이션 모델 중 2개(UNI·UNI2-h)에서만 5-seed 우연배제를 통과했고 Virchow2에서는 사전 기준을 통과하지 못하였다(real 0.9199 < 임계 0.9234, 마진 −0.0035 — 이는 셔플 null 산포가 넓은 것이며 신호 부재가 아니다; R5 참조). 또한 site 감사에서 site-label 구조화가 관찰되었다(Cramér's V = 0.378). HPV는 지도 한쪽 끝을 고정한 유일한 검정력 앵커일 뿐 법칙 일반화나 모델 무관 확증이 아니다. +† HPV 견고성은 3개 파운데이션 모델 중 2개(UNI·UNI2-h)에서만 5-seed 우연배제를 통과했고 Virchow2에서는 사전 기준을 통과하지 못하였다(real 0.9199 < 임계 0.9234, 마진 −0.0035 — 이는 셔플 null 산포가 넓은 것이며 신호 부재가 아니다; R5 참조). 또한 site 감사에서 site-label 구조화가 관찰되었다(Cramér's V = 0.397). HPV는 지도 한쪽 끝을 고정한 유일한 검정력 앵커일 뿐 법칙 일반화나 모델 무관 확증이 아니다. ‡ 폐 조직형(양성대조)은 site 감사에서 V(site, label) = 1.000이었다 — TCGA-LUAD/LUSC의 기관코드가 조직형과 100% 일치해 형태 신호와 site 서명이 분리 불가하다. 양성대조를 해석하는 모든 곳에 이 한계를 명시한다. ### R2. 우리 사전등록 분할에서는 변이축을 정할 수 없다 — 검정력 천장 @@ -166,7 +166,7 @@ Lauren diffuse는 원래 양성대조였다. 인환세포와 미만형은 강한 한계는 정면에 둔다. 모든 산출은 후향적·코호트 수준이며 `hypothesis_only`로, 개인 수준 benefit 주장이 아니다. -**Site/batch 교란과 염색 변이 한계 — "정말 형태인가?"라는 두 공격을 함께.** site-disjoint 분할은 동일 기관 슬라이드가 학습과 평가에 함께 들어가는 leakage를 방지했지만, 라벨과 기관의 결합(confounding)은 남아 있어 성능을 순수 형태 신호로 해석하는 데 한계가 있다. site/batch 교란 감사는 site-disjoint 분할이 라벨을 조직원천기관과 교란함을 다섯 개 endpoint에서 확인하였다 — 폐 조직형(양성대조) V = 1.000으로 형태와 site 서명이 분리 불가하고, 두경부 HPV V = 0.378, 위 Lauren은 유병률이 0.46에서 0.88로 이동하며, 폐 EGFR·KRAS도 site-label 연관이 유의하다. 이는 교란의 필요조건이지 "모델이 site를 읽는다"의 증명이 아니다. 별개로, H&E 염색 정규화는 주 파이프라인에 적용하지 않았다. 염색 정규화 견고성 점검은 유방 앵커 패턴(HER2 우연 수준, ER 높음, PAM50 보존)을 보존했지만, **이 점검은 유방 앵커에만 해당한다** — 다암종 헤드라인 축(두경부 HPV, 폐 조직형)은 raw 슬라이드가 소실되어 재추출이 보류돼 재점검하지 못하였다. 즉 리뷰어가 스캐너/염색 아티팩트로 가장 공격하기 쉬운 두 헤드라인 결과가 바로 site-교란 플래그를 달고 있으면서 아직 염색 검증이 안 된 축들이며, 이를 그대로 명시한다. H&E로부터의 site 예측성과 leave-one-site-out 성능이 교란 문제의 최종 판정을 하며, 그때까지 두 앵커를 해석하는 곳마다 이 한계를 명시한다. +**Site/batch 교란과 염색 변이 한계 — "정말 형태인가?"라는 두 공격을 함께.** site-disjoint 분할은 동일 기관 슬라이드가 학습과 평가에 함께 들어가는 leakage를 방지했지만, 라벨과 기관의 결합(confounding)은 남아 있어 성능을 순수 형태 신호로 해석하는 데 한계가 있다. site/batch 교란 감사는 site-disjoint 분할이 라벨을 조직원천기관과 교란함을 다섯 개 endpoint에서 확인하였다 — 폐 조직형(양성대조) V = 1.000으로 형태와 site 서명이 분리 불가하고, 두경부 HPV V = 0.397, 위 Lauren은 유병률이 0.46에서 0.88로 이동하며, 폐 EGFR·KRAS도 site-label 연관이 유의하다. 이는 교란의 필요조건이지 "모델이 site를 읽는다"의 증명이 아니다. 별개로, H&E 염색 정규화는 주 파이프라인에 적용하지 않았다. 염색 정규화 견고성 점검은 유방 앵커 패턴(HER2 우연 수준, ER 높음, PAM50 보존)을 보존했지만, **이 점검은 유방 앵커에만 해당한다** — 다암종 헤드라인 축(두경부 HPV, 폐 조직형)은 raw 슬라이드가 소실되어 재추출이 보류돼 재점검하지 못하였다. 즉 리뷰어가 스캐너/염색 아티팩트로 가장 공격하기 쉬운 두 헤드라인 결과가 바로 site-교란 플래그를 달고 있으면서 아직 염색 검증이 안 된 축들이며, 이를 그대로 명시한다. H&E로부터의 site 예측성과 leave-one-site-out 성능이 교란 문제의 최종 판정을 하며, 그때까지 두 앵커를 해석하는 곳마다 이 한계를 명시한다. 모델 비의존성에 대해서는, 폐 endpoint 간 상대적 AUROC 순서(조직형 > EGFR > KRAS)와 주요 음성 결과가 평가한 세 파운데이션 모델 전반에서 보존되었으나, 개별 분자 축의 우연배제 통과 여부는 모델에 따라 달랐다. 이는 순서 안정성이지 법칙 전체의 모델 비의존성이 아니다 — 헤드라인 단일 endpoint 확증(HPV, 대장 BRAF)은 각각 세 모델 중 둘에서만 우연배제를 통과했다. 20-seed 재점검은 Virchow2 HPV 임계가 안정화(≈0.837)되어 그 칸이 통과로 뒤집힐 것임을 시사했으나, **이 판정 변경은 채택하지 않는다** — braveji의 Critic 확인 대기 중이며(BIOP02-123) 보고 상태는 "3개 중 2개 모델"로 유지한다. diff --git a/manuscript/DRAFT_paperC_full_en.md b/manuscript/DRAFT_paperC_full_en.md index d244254..e422a52 100644 --- a/manuscript/DRAFT_paperC_full_en.md +++ b/manuscript/DRAFT_paperC_full_en.md @@ -61,7 +61,7 @@ Every endpoint is reported alongside a shuffle-null (5-seed), a prevalence basel | Breast | HER2 | Anchor (required axis) | 0.599 | near-random | Absent | **No signal supporting substitution · negative anchor** | | Gastric | Lauren diffuse | (Original positive control) | 0.536 (development 0.963) | pixel-mean 0.631 | Weakly present | Site-split case (R4) | -† HPV robustness passed 5-seed chance-exclusion in only 2 of 3 foundation models (UNI and UNI2-h) and failed in Virchow2 (real 0.9199 < threshold 0.9234; see R5). The site audit also found site-label structuring (Cramér's V = 0.378). HPV is the single powered anchor that fixes one end of the map — not a generalisation of the law and not a model-independent confirmation. +† HPV robustness passed 5-seed chance-exclusion in only 2 of 3 foundation models (UNI and UNI2-h) and failed in Virchow2 (real 0.9199 < threshold 0.9234; see R5). The site audit also found site-label structuring (Cramér's V = 0.397). HPV is the single powered anchor that fixes one end of the map — not a generalisation of the law and not a model-independent confirmation. ‡ For lung histology (positive control) the site audit gave V(site, label) = 1.000: TCGA-LUAD/LUSC institution codes coincide 100% with histology, so the morphological signal cannot be separated from the site signature. This limit is stated explicitly wherever the positive control is interpreted. ### R2. Our pre-registered split cannot decide the mutation axes — the power ceiling @@ -144,7 +144,7 @@ Identifying axes that are not legible in morphology is itself information the ma In breast HER2, routing from predicted subtype failed consistently in this cohort and routing definition, which indicates as cost a region where molecular testing remains necessary. We state the scheme dependence of per-axis cost honestly and restrict the robust claims to the anti-HER2 misassignment rate of 1.00 and the confidence interval of the contrast. Furthermore, when real treatment-outcome stratification was attached to the retrospective map, anti-HER2 pCR stratification was not significant (0.533), and this negative is consistent with the argument of the map. -The limits are placed in front. All results are retrospective, cohort-level and `hypothesis_only`; they are not claims of individual-level benefit. The site-disjoint split prevented leakage in which slides from the same institution enter both training and evaluation, but the coupling between label and institution (confounding) remains, which limits interpreting performance as a pure morphological signal; multi-institution generalisation remains subject to further prospective validation. In particular, a site/batch confounding audit confirmed in five endpoints that the site-disjoint split confounds labels with the tissue source site — lung histology (positive control) at V = 1.000, where morphology and site signature cannot be separated; head and neck HPV at V = 0.378; gastric Lauren, whose prevalence shifts from 0.46 to 0.88; and lung EGFR and KRAS, which also show significant site-label association. This is a necessary condition for confounding, not proof that "the model reads site". Site predictability from H&E and leave-one-site-out performance will adjudicate; until then this limit is stated wherever the two anchors are interpreted. +The limits are placed in front. All results are retrospective, cohort-level and `hypothesis_only`; they are not claims of individual-level benefit. The site-disjoint split prevented leakage in which slides from the same institution enter both training and evaluation, but the coupling between label and institution (confounding) remains, which limits interpreting performance as a pure morphological signal; multi-institution generalisation remains subject to further prospective validation. In particular, a site/batch confounding audit confirmed in five endpoints that the site-disjoint split confounds labels with the tissue source site — lung histology (positive control) at V = 1.000, where morphology and site signature cannot be separated; head and neck HPV at V = 0.397; gastric Lauren, whose prevalence shifts from 0.46 to 0.88; and lung EGFR and KRAS, which also show significant site-label association. This is a necessary condition for confounding, not proof that "the model reads site". Site predictability from H&E and leave-one-site-out performance will adjudicate; until then this limit is stated wherever the two anchors are interpreted. On model independence, the relative AUROC ordering among lung endpoints (histology > EGFR > KRAS) and the principal negative results were preserved across all three foundation models evaluated, whereas whether an individual molecular axis clears chance-exclusion varied by model. This is ordering stability, not model independence of the law as a whole, and no single cell of the map passed in all three models. diff --git a/manuscript/DRAFT_paperC_full_ko.md b/manuscript/DRAFT_paperC_full_ko.md index d351a30..d0a91a4 100644 --- a/manuscript/DRAFT_paperC_full_ko.md +++ b/manuscript/DRAFT_paperC_full_ko.md @@ -61,7 +61,7 @@ H&E 조직 이미지에서 종양의 분자 상태를 예측하는 연구는 이 | 유방 | HER2 | 앵커(필수축) | 0.599 | near-random | 없음 | **대체 지지 신호 없음 · 앵커 음성** | | 위 | Lauren diffuse | (원 양성대조) | 0.536 (개발 0.963) | pixel-mean 0.631 | 약하게 있음 | 기관분리 사례(R4) | -† HPV 견고성은 3개 파운데이션 모델 중 2개(UNI·UNI2-h)에서만 5-seed 우연배제를 통과했고 Virchow2에서는 통과하지 못하였다(real 0.9199 < 임계 0.9234; R5 참조). 또한 site 감사에서 site-label 구조화가 관찰되었다(Cramér's V = 0.378). HPV는 지도 한쪽 끝을 고정한 유일한 검정력 앵커일 뿐 법칙 일반화나 모델 무관 확증이 아니다. +† HPV 견고성은 3개 파운데이션 모델 중 2개(UNI·UNI2-h)에서만 5-seed 우연배제를 통과했고 Virchow2에서는 통과하지 못하였다(real 0.9199 < 임계 0.9234; R5 참조). 또한 site 감사에서 site-label 구조화가 관찰되었다(Cramér's V = 0.397). HPV는 지도 한쪽 끝을 고정한 유일한 검정력 앵커일 뿐 법칙 일반화나 모델 무관 확증이 아니다. ‡ 폐 조직형(양성대조)은 site 감사에서 V(site,label) = 1.000이었다 — TCGA-LUAD/LUSC의 기관코드가 조직형과 100% 일치해 형태 신호와 site 서명이 분리 불가하다. 양성대조로서의 해석에 이 한계를 명시한다. ### R2. 우리 사전등록 분할에서는 변이축을 정할 수 없다 — 검정력 천장 @@ -144,7 +144,7 @@ Lauren diffuse는 원래 양성대조였다. 인환세포와 미만형은 강한 유방 HER2에서는 예측 아형 라우팅이 이 코호트·라우팅 정의에서 일관되게 실패했고, 이는 분자검사가 필수인 영역을 비용으로 시사한다. 축별 비용의 스킴 의존성은 정직하게 서술하고, 견고한 주장은 항HER2 오배정률 1.00과 대비의 신뢰구간으로 한정한다. 나아가 후향적 지도에 실제 치료결과 층화를 달았을 때에도 항HER2 pCR 층화는 유의하지 않았고(0.533), 이 음성은 결정지도의 논지와 일관된다. -한계는 정면에 둔다. 모든 산출은 후향적·코호트 수준이며 `hypothesis_only`로, 개인 수준 benefit 주장이 아니다. site-disjoint 분할은 동일 기관 슬라이드가 학습과 평가에 함께 들어가는 leakage를 방지했지만, 라벨과 기관의 결합(confounding)은 남아 있어 성능을 순수 형태 신호로 해석하는 데 한계가 있으며, 다기관 일반화는 추가 전향 검증 대상이다. 특히 site/batch 교란 감사에서 site-disjoint 분할이 라벨을 조직원천기관과 교란함이 다섯 개 endpoint에서 확인되었다 — 폐 조직형(양성대조) V=1.000으로 형태와 site 서명이 분리 불가하고, 두경부 HPV V=0.378, 위 Lauren은 유병률이 0.46에서 0.88로 이동하며, 폐 EGFR·KRAS도 site-label 연관이 유의하다. 이는 교란의 필요조건이며 "모델이 site를 읽는다"의 증명은 아니다. H&E로부터의 site 예측성과 leave-one-site-out 성능이 최종 판정을 하며, 그때까지 양끝 앵커의 해석에 이 한계를 명시한다. +한계는 정면에 둔다. 모든 산출은 후향적·코호트 수준이며 `hypothesis_only`로, 개인 수준 benefit 주장이 아니다. site-disjoint 분할은 동일 기관 슬라이드가 학습과 평가에 함께 들어가는 leakage를 방지했지만, 라벨과 기관의 결합(confounding)은 남아 있어 성능을 순수 형태 신호로 해석하는 데 한계가 있으며, 다기관 일반화는 추가 전향 검증 대상이다. 특히 site/batch 교란 감사에서 site-disjoint 분할이 라벨을 조직원천기관과 교란함이 다섯 개 endpoint에서 확인되었다 — 폐 조직형(양성대조) V=1.000으로 형태와 site 서명이 분리 불가하고, 두경부 HPV V=0.397, 위 Lauren은 유병률이 0.46에서 0.88로 이동하며, 폐 EGFR·KRAS도 site-label 연관이 유의하다. 이는 교란의 필요조건이며 "모델이 site를 읽는다"의 증명은 아니다. H&E로부터의 site 예측성과 leave-one-site-out 성능이 최종 판정을 하며, 그때까지 양끝 앵커의 해석에 이 한계를 명시한다. 모델 비의존성에 대해서는, 폐 암종 내 endpoint 간 상대적 AUROC 순서(조직형 > EGFR > KRAS)와 주요 음성 결과가 평가한 세 파운데이션 모델 전반에서 보존되었으나, 개별 분자 축의 우연배제 통과 여부는 모델에 따라 다르게 나타났다. 이는 순서 안정성이지 법칙 전체의 모델 비의존성이 아니며, 단일 칸 확증이 세 모델 전부에서 통과한 것도 아니다. diff --git a/manuscript/DRAFT_paperC_sentence-skeleton_ko.md b/manuscript/DRAFT_paperC_sentence-skeleton_ko.md index 731d278..d2c4562 100644 --- a/manuscript/DRAFT_paperC_sentence-skeleton_ko.md +++ b/manuscript/DRAFT_paperC_sentence-skeleton_ko.md @@ -71,7 +71,7 @@ | 유방 | HER2 | 앵커(필수축) | 0.599 | near-random | 없음 | **대체 불가(앵커)** | | 위 | Lauren diffuse | (원 양성대조) | 0.536 (개발 0.963) | pixel-mean 0.631 | 약하게 있음 | 기관분리 사례(R4) | -† HPV 견고성은 3 FM 중 2개(UNI·UNI2-h)만 5-seed 우연배제, Virchow2 미통과(real 0.9199 < 임계 0.9234; R5 참조). 또한 #2 site 감사에서 site-label 구조화 V=0.378(BIOP02-123). HPV는 지도 한쪽 끝을 고정한 **유일한 검정력 앵커**일 뿐 법칙 일반화·모델 무관 확증이 아니다. +† HPV 견고성은 3 FM 중 2개(UNI·UNI2-h)만 5-seed 우연배제, Virchow2 미통과(real 0.9199 < 임계 0.9234; R5 참조). 또한 #2 site 감사에서 site-label 구조화 V=0.397(BIOP02-123). HPV는 지도 한쪽 끝을 고정한 **유일한 검정력 앵커**일 뿐 법칙 일반화·모델 무관 확증이 아니다. ‡ 폐 조직형(양성대조)은 #2 site 감사에서 V(site,label)=**1.000** — TCGA-LUAD/LUSC 기관코드가 조직형과 100% 일치해 형태 신호와 site 서명이 분리 불가(BIOP02-123). 양성대조로서의 해석에 이 한계를 명시한다. ### R2. 검정력 천장 — 우리 사전등록 분할에서는 변이축을 정할 수 없다 @@ -148,7 +148,7 @@ 3. (유방 HER2) 예측 아형 라우팅이 이 코호트·라우팅 정의에서 HER2축을 일관 실패시킴 = 분자검사 필수 영역을 비용으로 시사. per-axis cost의 스킴 의존성은 정직하게 서술하고 robust 주장은 항HER2 misroute 1.00 + contrast CI로 한정한다. 4. (실증 이빨 Yale) 후향적 지도에 실제 치료결과 층화를 달았고, 항HER2 pCR 층화 음성(0.533)은 결정지도 C 논지와 일관된다. 5. (한계) 전부 후향적·코호트 수준·hypothesis_only이며 개인 benefit이 아니다; site-disjoint로 leakage는 통제했으나 다기관 일반화는 추가 전향 검증 대상이다. - - **(site/batch 교란 — BIOP02-123 #2 감사 Phase 1)** site-disjoint 분할이 라벨을 조직원천기관(TSS)과 교란함이 5개 엔드포인트에서 확인됐다: 폐 조직형(양성대조) V=1.000(형태 vs site 서명 분리 불가), 두경부 HPV V=0.378, 위 Lauren 유병률 0.46→0.88(R4 정량 확증), 폐 EGFR·KRAS도 site-label 유의. 이는 교란의 **필요조건**이며 "모델이 site를 읽는다"의 증명은 아니다 — H&E→site 예측성·LOSO(Phase 2)가 최종 판정하며, 그때까지 양끝 앵커(HPV·폐 조직형) 해석에 이 한계를 명시한다. + - **(site/batch 교란 — BIOP02-123 #2 감사 Phase 1)** site-disjoint 분할이 라벨을 조직원천기관(TSS)과 교란함이 5개 엔드포인트에서 확인됐다: 폐 조직형(양성대조) V=1.000(형태 vs site 서명 분리 불가), 두경부 HPV V=0.397, 위 Lauren 유병률 0.46→0.88(R4 정량 확증), 폐 EGFR·KRAS도 site-label 유의. 이는 교란의 **필요조건**이며 "모델이 site를 읽는다"의 증명은 아니다 — H&E→site 예측성·LOSO(Phase 2)가 최종 판정하며, 그때까지 양끝 앵커(HPV·폐 조직형) 해석에 이 한계를 명시한다. 6. (모델 비의존성) 폐 결정지도의 축 순서와 주요 음성 결과가 세 FM에서 보존·재현됨을 보였다(Supplement); 이는 순서 안정성이지 법칙 전체의 모델 비의존성이 아니며, 단일 칸 확증이 세 모델 전부에서 통과한 것도 아니다. 7. (임상·연구 함의) 어디서 H&E 선별 후 표적 분자검사만 시행하면 비용을 아끼고 어디서 대체가 위험한지에 대한 전향 검증 설계용 의사결정 틀·우선순위 지도이며, 가치는 비싸거나 느리거나 희소한 분자검사·자원제한 세팅에 집중된다(임상 권고·전면 대체 주장 아님). 8. (골드스탠다드 방어) 기여는 "골드스탠다드를 이긴다"가 아니라 "값싼 H&E가 언제 분자검사를 pre-screen/triage할 수 있고 언제 불가한지의 예측 가능한 결정지도"다. diff --git a/manuscript/LONG_ABSTRACT_GIW2026_BIOP02_en.md b/manuscript/LONG_ABSTRACT_GIW2026_BIOP02_en.md index 7c944ec..b9ff674 100644 --- a/manuscript/LONG_ABSTRACT_GIW2026_BIOP02_en.md +++ b/manuscript/LONG_ABSTRACT_GIW2026_BIOP02_en.md @@ -41,7 +41,7 @@ Of roughly 15 endpoints, **exactly one non-control endpoint met the pre-register | Lung KRAS-G12C | 0.681 | [0.577–0.783] | 14 | Undecided | | Gastric ERBB2 amplification | 0.644 | [0.523–0.771] | 14 | No signal | -† The head and neck HPV confirmation holds under the primary model (UNI). Of 3 foundation models, UNI and UNI2-h cleared the null but Virchow2 did not (0.9199 against a threshold of 0.9234). The site audit also found an association between label and tissue source site (Cramér's V = 0.378). +† The head and neck HPV confirmation holds under the primary model (UNI). Of 3 foundation models, UNI and UNI2-h cleared the null but Virchow2 did not (0.9199 against a threshold of 0.9234). The site audit also found an association between label and tissue source site (Cramér's V = 0.397). **Table 2. Anchor and retrospective cohorts — breast and colorectal (not counted).** diff --git a/manuscript/LONG_ABSTRACT_GIW2026_BIOP02_ko.md b/manuscript/LONG_ABSTRACT_GIW2026_BIOP02_ko.md index c1e192a..1c989ad 100644 --- a/manuscript/LONG_ABSTRACT_GIW2026_BIOP02_ko.md +++ b/manuscript/LONG_ABSTRACT_GIW2026_BIOP02_ko.md @@ -41,7 +41,7 @@ AI를 활용한 H&E 이미지 분석에서 분자 아형을 예측하는 연구 | 폐암 KRAS-G12C | 0.681 | [0.577–0.783] | 14 | 판정 보류 | | 위암 ERBB2 증폭 | 0.644 | [0.523–0.771] | 14 | 신호 없음 | -※ 두경부암 HPV 확증은 주모델(UNI)에서 성립한다. 파운데이션 모델 3종 중 UNI·UNI2-h는 귀무분포를 넘었으나 Virchow2는 넘지 못하였다(0.9199, 임계 0.9234). 기관 교란 감사에서 라벨과 조직원천기관의 연관도 확인되었다(Cramér's V = 0.378). +※ 두경부암 HPV 확증은 주모델(UNI)에서 성립한다. 파운데이션 모델 3종 중 UNI·UNI2-h는 귀무분포를 넘었으나 Virchow2는 넘지 못하였다(0.9199, 임계 0.9234). 기관 교란 감사에서 라벨과 조직원천기관의 연관도 확인되었다(Cramér's V = 0.397). **표 2. 앵커 및 회고적 코호트 — 유방암·대장암 (집계 제외).**