medsci-skills 5.27.0 → 5.28.0

package/metadata/distribution_files.json

@@ -183,8 +183,8 @@
     },
     {
       "path": "skills/analyze-stats/SKILL.md",
-      "size": 60456,
-      "sha256": "c541c8811b918bb5b5cb380297a52659b4bda88537841574dc1272a8f7918aa8"
+      "size": 60966,
+      "sha256": "dca033657405001ba3e6f421a258edd1c9cc17c62e7661d0f987109cd4db33cf"
     },
     {
       "path": "skills/analyze-stats/references/analysis_guides/agreement_reliability.md",
@@ -338,8 +338,8 @@
     },
     {
       "path": "skills/analyze-stats/references/table-standards/table-types/incremental_value.md",
-      "size": 4488,
-      "sha256": "7a79208ff425c306e877b0938dbe8e1c8f6e663d32f5509496ef1ff87ff18b04"
+      "size": 8288,
+      "sha256": "8cd2c321a7e902e730ad77f30e45be95e489bd9344171e176c558009f8bec153"
     },
     {
       "path": "skills/analyze-stats/references/table-standards/table-types/meta_analysis.md",
@@ -1083,8 +1083,8 @@
     },
     {
       "path": "skills/deidentify/deidentify.py",
-      "size": 46012,
-      "sha256": "9aa00c186b7bf9d9130be5a79af8645004be40ab12ec3ae9171a632da67b4cb9"
+      "size": 46778,
+      "sha256": "5931af2083887f202842f7e1a9df79a501436c004f749ce1e8d833d7d814601a"
     },
     {
       "path": "skills/deidentify/locales/_template.json",
@@ -1456,6 +1456,11 @@
       "size": 1829,
       "sha256": "2d6a19ecd70d20211a31ab0be294793dea4b0ec1c0410f6e2147cb2022c7a3a6"
     },
+    {
+      "path": "skills/find-journal/references/journal_profiles/Annals_of_Oncology.md",
+      "size": 3586,
+      "sha256": "33a169e9458616cf3261019e3ffe8cd699901ab8fa2706cc4298b035b5edffe3"
+    },
     {
       "path": "skills/find-journal/references/journal_profiles/Artificial_Intelligence_in_Medicine.md",
       "size": 1554,
@@ -1463,8 +1468,8 @@
     },
     {
       "path": "skills/find-journal/references/journal_profiles/BJR_Case_Reports.md",
-      "size": 2044,
-      "sha256": "74eac3cbdacfc9fddbcb6b6070e2be91b37a8042f9e2e4f34a4c2eb25a3041bf"
+      "size": 2041,
+      "sha256": "579704a717ee7e79cd9c2ca2b1c55ac25af33b9a10cd87442f72026b073885d3"
     },
     {
       "path": "skills/find-journal/references/journal_profiles/BMC_Medicine.md",
@@ -1473,8 +1478,8 @@
     },
     {
       "path": "skills/find-journal/references/journal_profiles/BMJ_Case_Reports.md",
-      "size": 2145,
-      "sha256": "029e4d1a1860f6d93345278726c0195d1c0a7ceaed1fe7510441c89efd953f7d"
+      "size": 2142,
+      "sha256": "7662c66eb32bc6c40815d64ae08f923ab6ba8506725988bdb5db95a1f3a13601"
     },
     {
       "path": "skills/find-journal/references/journal_profiles/British_Journal_of_Radiology.md",
@@ -1483,13 +1488,23 @@
     },
     {
       "path": "skills/find-journal/references/journal_profiles/CVIR.md",
-      "size": 2881,
-      "sha256": "a28b761930d158def0153408ea8d152c1a27d7122c2043178509405090efad1f"
+      "size": 2878,
+      "sha256": "78016adb4143aca41e8861751896bf780da66f61fceb6303ab06996190abe83e"
+    },
+    {
+      "path": "skills/find-journal/references/journal_profiles/Cancer_Research_and_Treatment.md",
+      "size": 4032,
+      "sha256": "db49c78c140d5d6a49d6fecd013365ee0a72b43ed97a551b4fd9be0e2a41f51d"
     },
     {
       "path": "skills/find-journal/references/journal_profiles/Chest.md",
-      "size": 3386,
-      "sha256": "99f7780968ba1457dda311c48e6613b34d752f0165665094fc516a8258d06801"
+      "size": 3383,
+      "sha256": "4a6f33f8167caa195f885d76ef299dd1a9be00fda979ffd8c3595bbd88c17a62"
+    },
+    {
+      "path": "skills/find-journal/references/journal_profiles/Clinical_Cancer_Research.md",
+      "size": 4381,
+      "sha256": "44123275b0643186a63ed0bd5e184676da5f5d8d53e7fba27f791f9d885b5560"
     },
     {
       "path": "skills/find-journal/references/journal_profiles/Clinical_Radiology.md",
@@ -1503,8 +1518,8 @@
     },
     {
       "path": "skills/find-journal/references/journal_profiles/Cureus.md",
-      "size": 1836,
-      "sha256": "daafc2d6ec1e2058a9d6ae037c9ed29fc5ef200695a52016f846e80ced91ab2f"
+      "size": 1833,
+      "sha256": "d7c20f59e305d6626fd0a79fad94e67837aeff14c968a487afb75e8325efe01e"
     },
     {
       "path": "skills/find-journal/references/journal_profiles/Diabetes_Metabolism_Journal.md",
@@ -1516,6 +1531,11 @@
       "size": 2210,
       "sha256": "ea2ba5ab54862bea842c799614c3ed16617d50ff40178d3d721fbd6af6541aef"
     },
+    {
+      "path": "skills/find-journal/references/journal_profiles/ESMO_Open.md",
+      "size": 4030,
+      "sha256": "ff8ed0deb9dcb405fad3de565f964e3fccbf3e1fc68dda22812e1edd6c7e0291"
+    },
     {
       "path": "skills/find-journal/references/journal_profiles/Endocrinology_and_Metabolism.md",
       "size": 2672,
@@ -1626,11 +1646,21 @@
       "size": 1759,
       "sha256": "b0a2a7aa791e86b375ba19e63036106ab712cb14f4c9ca0398de75dc19e3f946"
     },
+    {
+      "path": "skills/find-journal/references/journal_profiles/Journal_of_Breast_Cancer.md",
+      "size": 3517,
+      "sha256": "66d2fd34f44ca75e83008543886532c7dea2a8c89137d7851b7780f048cab52b"
+    },
     {
       "path": "skills/find-journal/references/journal_profiles/Journal_of_Clinical_Endocrinology_and_Metabolism.md",
       "size": 2837,
       "sha256": "8c40a36ec7409b86d4e222a71b4bb39deb453d038ca83da1d7761ade2517657f"
     },
+    {
+      "path": "skills/find-journal/references/journal_profiles/Journal_of_Clinical_Oncology.md",
+      "size": 4268,
+      "sha256": "ce289ff9a5b5ed1317855431ecb08d5cfb103e05c5b0825924a36c7b78568d86"
+    },
     {
       "path": "skills/find-journal/references/journal_profiles/Journal_of_Magnetic_Resonance_Imaging.md",
       "size": 1576,
@@ -1638,8 +1668,8 @@
     },
     {
       "path": "skills/find-journal/references/journal_profiles/Journal_of_Medical_Case_Reports.md",
-      "size": 1954,
-      "sha256": "60faecafdc259db288cf91fb1430eeb619f47e73886c1dad483c05ae9d54f22d"
+      "size": 1951,
+      "sha256": "149ccfb0c2dfc482169c5897c9e5cab6a44f7f71e827cf1be169db6ba3486dee"
     },
     {
       "path": "skills/find-journal/references/journal_profiles/Journal_of_Nuclear_Medicine.md",
@@ -1768,8 +1798,8 @@
     },
     {
       "path": "skills/find-journal/references/journal_profiles/Radiology_Case_Reports.md",
-      "size": 1849,
-      "sha256": "87067ec91a2d6c2d3d4e8dd00f41e0ca065926f4db36f0a20f273b6f0d2c2925"
+      "size": 1846,
+      "sha256": "4fd5ec1cc30c9d87b9c539bddf3c73ea455e8255069e92c00ec07559d94b157b"
     },
     {
       "path": "skills/find-journal/references/journal_profiles/Skeletal_Radiology.md",
@@ -1793,14 +1823,19 @@
     },
     {
       "path": "skills/find-journal/references/journal_profiles/The_Lancet_Digital_Health.md",
-      "size": 2234,
-      "sha256": "35bb895cadb1867b081184cc7b044f7c95dc7b6c9289a3ce2d311e9eec811214"
+      "size": 2231,
+      "sha256": "bb3ac6c10aa733cb4fbeb823ed5288001fa5a72e8aa0e47cd87b9f80a0b0310d"
     },
     {
       "path": "skills/find-journal/references/journal_profiles/World_Journal_of_Hepatology.md",
       "size": 3707,
       "sha256": "912a14910717f00306a8f94a7b1ec4cd81a2690dae0dcabe1a85d7e82e109eb2"
     },
+    {
+      "path": "skills/find-journal/references/journal_profiles/npj_Breast_Cancer.md",
+      "size": 4589,
+      "sha256": "af2714dd8f07b12bfa9b54bdd1c711781271c0727ace6cee87ce763d553908e8"
+    },
     {
       "path": "skills/find-journal/references/journal_profiles/npj_Digital_Medicine.md",
       "size": 1511,
@@ -2198,8 +2233,8 @@
     },
     {
       "path": "skills/make-figures/references/exemplar_plots/README.md",
-      "size": 10492,
-      "sha256": "8840d2b36492f10afc4f17340903c8c052fa78f0966a3379d401996830c41f70"
+      "size": 11082,
+      "sha256": "55d62362fa4d5a3a08899d869739a2d81ceae9c53a1f42ef84d463caf5f70be6"
     },
     {
       "path": "skills/make-figures/references/exemplar_plots/ai_assistance_effect.md",
@@ -2281,6 +2316,11 @@
       "size": 4581,
       "sha256": "9bedf2f0b762702b1b9f794d6e83abc512a0bd504101ac27133558074a788b04"
     },
+    {
+      "path": "skills/make-figures/references/exemplar_plots/shap_beeswarm.md",
+      "size": 4850,
+      "sha256": "bc3de2226c96d4d935d5291b4058e0cdd1dc03e807004391347b3ba32a6311d5"
+    },
     {
       "path": "skills/make-figures/references/exemplar_plots/spider_plot.md",
       "size": 3114,
@@ -6131,6 +6171,11 @@
       "size": 6387,
       "sha256": "58e6ee1186985e75054194070a8e97e871fe6207bfaab07264d8ea1e5cc330eb"
     },
+    {
+      "path": "skills/write-paper/references/journal_profiles/Annals_of_Oncology.md",
+      "size": 7859,
+      "sha256": "5352cfac5c6a396c78fc506224371b5ee6fc69bdd26694d261ef299cf1dfab09"
+    },
     {
       "path": "skills/write-paper/references/journal_profiles/Artificial_Intelligence_in_Medicine.md",
       "size": 3492,
@@ -6146,10 +6191,20 @@
       "size": 7825,
       "sha256": "db438eb8e86a2f6f6da37c9495eaf65d45bb6139f42b409a3c3a16ca1da78ea8"
     },
+    {
+      "path": "skills/write-paper/references/journal_profiles/Cancer_Research_and_Treatment.md",
+      "size": 7484,
+      "sha256": "c1549dec158b9e00a3032c4da78a61ad3a0919f736e2f7d24d07af0afc02e94b"
+    },
     {
       "path": "skills/write-paper/references/journal_profiles/Chest.md",
-      "size": 16319,
-      "sha256": "092cd7c888e1075abc13b59941de604b21201addf27301751db1b28ba975cd21"
+      "size": 16316,
+      "sha256": "24f7de1e22a61aee85b22c0b34cbf13a237ca04065a14512102c83b9ed584808"
+    },
+    {
+      "path": "skills/write-paper/references/journal_profiles/Clinical_Cancer_Research.md",
+      "size": 10002,
+      "sha256": "d45c8aa63feb2ad197a19bffdcef477f7ac801c374a7a30c13591c2ffff7d2c4"
     },
     {
       "path": "skills/write-paper/references/journal_profiles/Clinical_Radiology.md",
@@ -6171,6 +6226,11 @@
       "size": 12383,
       "sha256": "3ebcba0749b4a70c4f7b0d2dc1ec09cc28b982f942d879db9719808315829c0e"
     },
+    {
+      "path": "skills/write-paper/references/journal_profiles/ESMO_Open.md",
+      "size": 8040,
+      "sha256": "2c4d46e246d4a8f2cc1bff1a7d3a4c422ad66da1bdb46e01ef08828dd93e20e6"
+    },
     {
       "path": "skills/write-paper/references/journal_profiles/Endocrinology_and_Metabolism.md",
       "size": 7164,
@@ -6263,23 +6323,33 @@
     },
     {
       "path": "skills/write-paper/references/journal_profiles/JNIS.md",
-      "size": 10421,
-      "sha256": "6314ce1c3c57810d9a48dd9eb277d473f402cad063ba1c1d67ca811ee9d75e09"
+      "size": 10418,
+      "sha256": "0ba37375a08032323ee15dae79c878957e2099b50b8ebf26d7caa4e1e9e50df4"
     },
     {
       "path": "skills/write-paper/references/journal_profiles/JVIR.md",
       "size": 6124,
       "sha256": "313604759c6e6471ec82e55ab0db489058fe2a7a2628f4fae92471423fb64f60"
     },
+    {
+      "path": "skills/write-paper/references/journal_profiles/Journal_of_Breast_Cancer.md",
+      "size": 7277,
+      "sha256": "0dc4bede6e74c540ce5b8326db5d49b245f76867e5951703abc2ad07c0c515af"
+    },
     {
       "path": "skills/write-paper/references/journal_profiles/Journal_of_Clinical_Endocrinology_and_Metabolism.md",
       "size": 9600,
       "sha256": "2cade69fc364915be5ff06cee7a6b77b9bde0c5499fac75f05007332affca8d8"
     },
+    {
+      "path": "skills/write-paper/references/journal_profiles/Journal_of_Clinical_Oncology.md",
+      "size": 9172,
+      "sha256": "582dc60183235ed628c0c4483bcc803e337ea62e8fdd5856ccfc2cd1277dbc41"
+    },
     {
       "path": "skills/write-paper/references/journal_profiles/Journal_of_Stroke.md",
-      "size": 7740,
-      "sha256": "e551e78c799c6a26754f3186ee6d7ce0dd6931e5dc352dc225ee30e94cf1dca7"
+      "size": 7737,
+      "sha256": "079a49839844824645be89c0d5abbccace3ddb765f925f2bee0ec43cc8c806c3"
     },
     {
       "path": "skills/write-paper/references/journal_profiles/KJR.md",
@@ -6308,8 +6378,8 @@
     },
     {
       "path": "skills/write-paper/references/journal_profiles/Liver_International.md",
-      "size": 12373,
-      "sha256": "f6ff23ae9ca9b7d6cbc6d3250a6098c45c4e1c9c36f263a77647acac9a2bcadd"
+      "size": 12370,
+      "sha256": "e0d280ab2302a09f03a4be6fea6c4518f48ee1c0b9558cf8158006223a6e6fa2"
     },
     {
       "path": "skills/write-paper/references/journal_profiles/Medical_Image_Analysis.md",
@@ -6323,8 +6393,8 @@
     },
     {
       "path": "skills/write-paper/references/journal_profiles/NEJM_AI.md",
-      "size": 9486,
-      "sha256": "6155b1377e626d42b59b957d1752a0ff226731e5d57dee14e7ab1a1ed346665b"
+      "size": 9483,
+      "sha256": "2c3880cd309eaacbda6b1a82e85bb4efea4735b03e8fe4b2feb27a4e817d2b8e"
     },
     {
       "path": "skills/write-paper/references/journal_profiles/Nature_Medicine.md",
@@ -6363,8 +6433,8 @@
     },
     {
       "path": "skills/write-paper/references/journal_profiles/Stroke.md",
-      "size": 8786,
-      "sha256": "c4294b7b897932c9b3a9c700da39fb770814acdda3b2b9471ca81f1b773b5022"
+      "size": 8780,
+      "sha256": "6c7ab7b6fd0ee1c18f905333bd42770f6e8b430eab7edb692a7b3cf9ba775df4"
     },
     {
       "path": "skills/write-paper/references/journal_profiles/The_BMJ.md",
@@ -6378,14 +6448,19 @@
     },
     {
       "path": "skills/write-paper/references/journal_profiles/The_Lancet_Digital_Health.md",
-      "size": 8947,
-      "sha256": "5651fca14c1b2d78cb0b535d431487a6904f0e18c354dab52e192cccb8dc9014"
+      "size": 8944,
+      "sha256": "723b0f5e91be86298401ea2f1b6e76675881675de1a7622dde8de906f0033a2f"
     },
     {
       "path": "skills/write-paper/references/journal_profiles/World_Journal_of_Hepatology.md",
       "size": 7289,
       "sha256": "4f8594e07080dafb56dc6945addf45e7ceddbc3185f781f4edcd211744a8fb24"
     },
+    {
+      "path": "skills/write-paper/references/journal_profiles/npj_Breast_Cancer.md",
+      "size": 10165,
+      "sha256": "84eced761fe6094146bde8f09f812b1df28c464997c07c36c641c076a170f832"
+    },
     {
       "path": "skills/write-paper/references/journal_profiles/npj_Digital_Medicine.md",
       "size": 6000,

package/metadata/distribution_manifest.json

@@ -1,6 +1,6 @@
 {
   "schema_version": 1,
-  "version": "5.27.0",
+  "version": "5.28.0",
   "owned_skills": [
     "academic-aio",
     "add-journal",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "medsci-skills",
-  "version": "5.27.0",
+  "version": "5.28.0",
   "description": "MedSci Skills — a medical/scientific research skill suite for AI coding agents (Claude Code, Codex, Cursor, Copilot). The npm package is a terminal-friendly installer shortcut; the canonical distribution remains the GitHub repository and the Claude Code plugin marketplace.",
   "license": "SEE LICENSE IN LICENSE",
   "homepage": "https://github.com/Aperivue/medsci-skills#readme",

package/skills/analyze-stats/SKILL.md

@@ -415,16 +415,16 @@ tbl %>% as_flex_table() %>% flextable::save_as_docx(path = "table.docx")
 - Always report: sensitivity, specificity, PPV, NPV, accuracy, AUC
 - CIs: Wilson score for proportions, DeLong for AUC
 - ROC curve: include diagonal reference line, AUC in legend
-- If comparing models: DeLong test for AUC comparison
+- If comparing models: paired DeLong test for two separate scores on the same patients. For a nested "baseline + new term" model fitted and evaluated on the same data, do not use it as the added-value test; test the term (likelihood ratio) and report ΔAUC with its CI (see the table guide below)
 - Youden's index for optimal threshold when applicable
 - Include calibration assessment (Brier score, calibration plot) for prediction models
-- **NRI/IDI**: When comparing two models (e.g., base model vs model + AI score), report:
-  - Category-based NRI (with clinically defined risk categories)
-  - Continuous NRI (note: tends to be inflated — report alongside category-based)
+- **NRI/IDI**: If reported when comparing two models (e.g., base model vs model + AI score):
+  - NRI separately for events and non-events (with two risk categories these are the changes in sensitivity and specificity)
+  - Categorical NRI only with prespecified, clinically defined risk categories; never the continuous (category-free) NRI as evidence of added value — it can overstate it even in validation data
   - IDI (Integrated Discrimination Improvement)
   - Bootstrap 95% CIs (1000+ iterations)
-  - These supplement, not replace, DeLong AUC comparison
-- Table type guide (added value beyond a baseline): `references/table-standards/table-types/incremental_value.md` (paired ΔAUC + DeLong CI, continuous NRI with event/non-event split, IDI, net benefit at a prespecified threshold, same-patient/calibrated-first discipline). Pairs the decision-curve exemplar `make-figures` `references/exemplar_plots/decision_curve.md`.
+  - These supplement the added-value test and ΔAUC with its CI; net benefit is the preferred single summary
+- Table type guide (added value beyond a baseline): `references/table-standards/table-types/incremental_value.md` (likelihood-ratio test of the new term, paired ΔAUC with CI as an estimate, NRI split by events/non-events, IDI, net benefit at a prespecified threshold, same-patient/calibrated-first discipline). Pairs the decision-curve exemplar `make-figures` `references/exemplar_plots/decision_curve.md`.
 - Reader study (MRMC): `references/table-standards/table-types/reader_study.md` (per-reader + reader-averaged AUC with an Obuchowski–Rockette/DBM reader+case CI, per-patient vs per-lesion unit, superiority vs non-inferiority margin). Use an MRMC method (not a fixed-reader DeLong CI) for a claim that generalises to readers. Pairs `make-figures` `references/exemplar_plots/mrmc_roc.md`.
 
 ### Inter-rater Agreement

package/skills/analyze-stats/references/table-standards/table-types/incremental_value.md

@@ -3,8 +3,9 @@
 The table standard for any "the new marker/model adds value **beyond** an established baseline"
 claim — a new biomarker on top of a clinical score, an AI model on top of a radiologist or a
 guideline nomogram. Discrimination alone (a higher AUROC) is not sufficient evidence of added
-value; report the paired change in discrimination **with** a reclassification and a clinical-utility
-metric, all on the **same patients** and the same held-out data.
+value; report the paired change in discrimination **with** a clinical-utility metric (and
+reclassification, if used, reported as below), all on the **same patients** and the same held-out
+data.
 
 ## Reporting Guidelines
 - **TRIPOD / TRIPOD+AI**: prediction-model development/validation, model comparison and performance
@@ -17,13 +18,15 @@ metric, all on the **same patients** and the same held-out data.
 ```
 Table 4. Incremental Value of [New Predictor/Model] Added to [Baseline] (external test set, n = ___, events = ___)
 
-Model                         C-statistic (95% CI)   ΔC (95% CI)         P (ΔC)   Continuous NRI (95% CI)   IDI (95% CI)      ΔNet benefit @ threshold
-Baseline ([predictors])       0.74 (0.70-0.78)       —  (reference)       —        —                          —                 — (reference)
-Baseline + [new predictor]    0.81 (0.77-0.85)       0.07 (0.04-0.10)     <.001    0.38 (0.21-0.55)           0.045 (0.02-0.07) +0.021 @ 10%
+Model                         C-statistic (95% CI)   ΔC (95% CI)         NRI events / non-events (95% CI)   IDI (95% CI)      ΔNet benefit @ threshold (95% CI)
+Baseline ([predictors])       0.74 (0.70-0.78)       —  (reference)       —                                  —                 — (reference)
+Baseline + [new predictor]    0.81 (0.77-0.85)       0.07 (0.04-0.10)     ___ / ___                          0.045 (0.02-0.07) +0.021 (___-___) @ 10%
 
-ΔC by DeLong test on paired, same-patient predictions. NRI/IDI from the paired risk
-estimates (event/non-event components reported separately in the footnote). Net benefit
-from decision-curve analysis at the prespecified threshold; full curve in Figure X.
+Added-value test: likelihood-ratio test of [new predictor] in the development cohort
+(n = ___, events = ___), χ²(df) = ___, P = ___. ΔC is the size of the gain on these held-out
+patients with a paired 95% CI (DeLong or bootstrap) — an estimate, not a second test. NRI/IDI
+from the paired risk estimates, NRI given separately for events and non-events. Net benefit from
+decision-curve analysis at the prespecified threshold; full curve in Figure X.
 ```
 
 ## Rules
@@ -31,41 +34,82 @@ from decision-curve analysis at the prespecified threshold; full curve in Figure
   term, evaluated on the **identical** held-out patients — not two models on different cohorts.
 - **Baseline named and justified**: state exactly what the baseline contains (clinical score,
   guideline nomogram, prior model) and that it was applied as published / recalibrated / refit.
-- **ΔAUC with a paired CI and test**: report ΔC-statistic with a 95% CI and the **DeLong** paired
-  test, not two independent AUROCs eyeballed side by side. A tiny, non-significant ΔC with wide CI
-  is not "added value."
-- **Reclassification reported correctly**: prefer **continuous NRI** or category-free NRI; if a
-  categorical NRI is used, prespecify and justify the risk categories (post-hoc categories inflate
-  NRI). Always report **event** and **non-event** NRI components separately — a positive overall
-  NRI driven only by the non-event component is weak. Report **IDI** alongside.
+- **One test for "adds anything", then an estimate for "how much"**: for a prespecified term added
+  to a regression baseline, whether it adds information is tested through its coefficient in the
+  full model. That is the **likelihood-ratio** (or Wald) test in the development data, with both
+  models fitted on the same observations. ΔC-statistic is then reported as the **size** of the
+  gain, with a paired 95% CI, not as a second significance test. A DeLong test of ΔAUC between
+  nested models **fitted and evaluated on the same data** is the wrong test for added value. When
+  the new term has no association, the statistic's null distribution is not normal, and the test
+  is conservative for small and moderate effects (Demler, Pencina & D'Agostino, *Stat Med* 2012).
+  In simulation its size fell below 0.006, with much lower power than the likelihood-ratio and
+  Wald tests (Vickers, Cronin & Begg, *BMC Med Res Methodol* 2011). A marker with a modest real
+  association can therefore "fail" it. On held-out data, comparing two models **frozen before
+  evaluation** is what the DeLong method was built for: correlated ROC curves of given scores on
+  the same patients. In simulation it held its nominal type I error there (Chen et al., *BMC Med
+  Res Methodol* 2013). It then answers whether those two scores differ in AUC, not whether the
+  new term carries information. When the augmented model is not a regression with a separable
+  prespecified term, such as a retrained network or a data-selected predictor set, that held-out
+  paired comparison is the evidence, not a coefficient test. Never compare two independent AUROCs
+  by eye, side by side. A tiny ΔC with a wide CI is not evidence of added discrimination, though on
+  its own it does not rule out a gain in calibration or net benefit.
+- **Reclassification, if reported, reported correctly**: give NRI **separately for events and
+  non-events**. An overall NRI can be positive from one side alone, and which side moved decides
+  whether the change matters clinically: more events caught, or fewer non-events flagged. With two
+  risk categories those components are simply the changes in sensitivity and specificity, so report
+  them under those names. **Do not present the category-free (continuous) NRI as evidence of added
+  value**: it can overstate a marker's incremental value even in independent validation data. With
+  three or more categories the same review recommends against NRI altogether (Kerr et al.,
+  *Epidemiology* 2014). If a categorical NRI is used anyway, prespecify and justify the categories
+  (post-hoc categories inflate NRI), and bootstrap its CI rather than using published variance
+  formulas. IDI, if reported, goes alongside with its CI.
 - **Clinical utility, not just statistics**: include **net benefit** (decision-curve analysis) at a
   prespecified, clinically justified threshold (and reference the full curve —
-  `make-figures` `exemplar_plots/decision_curve.md`). Reclassification metrics without utility can
-  mislead.
-- **Calibration first**: NRI/IDI/net benefit all depend on predicted probabilities, so both models
-  must be calibrated on the test data before these are computed (pair with the calibration plot).
+  `make-figures` `exemplar_plots/decision_curve.md`). The improvement in net benefit is the
+  preferred single-number summary of the increment (Kerr et al. 2014). Reclassification metrics
+  without utility can mislead.
+- **Calibration first**: NRI/IDI/net benefit all depend on predicted probabilities, so assess both
+  models' calibration on the test data before computing them (pair with the calibration plot). If
+  a model has to be recalibrated, do it on data other than the test set. Recalibrating on the test
+  outcomes and then evaluating on them is no longer a validation of a frozen model.
 - **One decimal discipline / CIs everywhere**: every estimate carries a 95% CI; bootstrap CIs state
   the number of replicates; the same bootstrap resamples are reused across paired metrics.
 
 ## Common pitfalls (flag in review)
-- Reporting only ΔAUROC and calling it "added value" (no reclassification, no utility).
+- Reporting only ΔAUROC and calling it "added value" (no clinical-utility metric).
 - NRI with post-hoc categories, or only the overall NRI without event/non-event split.
+- A DeLong test of ΔAUC used as the significance test for an added predictor in nested models
+  fitted and evaluated on the same data. It is conservative and underpowered there, so test the
+  coefficient.
+- A continuous (category-free) NRI offered as the evidence of added value.
 - New vs baseline AUROCs from different cohorts or different n (not a paired comparison).
 - Net benefit asserted in prose with no decision curve and no stated threshold.
 
 ## Code (illustrative)
 
 ```r
-# Paired ΔAUC with DeLong CI (pROC); NRI/IDI (Hmisc / nricens / PredictABEL-style).
+# 1) Does the new term add anything? One test: likelihood ratio, in the development data.
+#    Fit both models on the same complete cases, or the comparison is not nested.
+dev <- dev[complete.cases(dev[, c("y", "x1", "x2", "marker")]), ]
+fit_base <- glm(y ~ x1 + x2,          family = binomial, data = dev)
+fit_full <- glm(y ~ x1 + x2 + marker, family = binomial, data = dev)
+anova(fit_base, fit_full, test = "LRT")        # Cox: anova(cox_base, cox_full)
+# 2) How much? Paired ΔAUC with its CI on the held-out patients -- an estimate, not a 2nd test.
+#    Fix the outcome levels and direction so roc() cannot flip a score on the test data.
 library(pROC)
-roc_base <- roc(y, p_baseline); roc_full <- roc(y, p_full)
-roc.test(roc_base, roc_full, method = "delong", paired = TRUE)   # ΔC + p, paired
-# NRI/IDI from paired risk vectors p_baseline, p_full (report event + non-event components);
+roc_base <- roc(y_test, p_baseline, levels = c(0, 1), direction = "<")
+roc_full <- roc(y_test, p_full,     levels = c(0, 1), direction = "<")
+roc.test(roc_full, roc_base, method = "delong", paired = TRUE)$conf.int   # CI of full - base
+# NRI (events / non-events separately) and IDI from the paired risk vectors p_baseline, p_full;
 # net benefit via the decision-curve template (analyze-stats references/templates/dca_plot.R).
 ```
 
 ```python
-# Paired DeLong CI for ΔAUC (e.g., delong_roc_test); reclassification via statsmodels/lifelines-
-# adjacent helpers. Compute NRI/IDI on the SAME held-out patients used for the AUCs; reuse one set
-# of bootstrap resamples across ΔAUC, NRI, IDI, and net benefit so the CIs are mutually consistent.
+# 1) One test for "adds anything": likelihood ratio of the nested fits (development data),
+#    both fitted on the same observations.
+from scipy.stats import chi2
+lr = 2 * (fit_full.llf - fit_base.llf)          # statsmodels Logit / GLM results
+p = chi2.sf(lr, fit_full.df_model - fit_base.df_model)
+# 2) ΔAUC on the SAME held-out patients with a paired bootstrap CI (an estimate, not a 2nd test);
+# reuse one set of resamples across ΔAUC, NRI, IDI and net benefit so the CIs are consistent.
 ```

package/skills/deidentify/deidentify.py

@@ -163,11 +163,14 @@ def build_locale_patterns(locale: dict) -> tuple[
     # Value patterns: universal + locale-specific
     value_patterns = list(UNIVERSAL_VALUE_PATTERNS)
 
-    # National ID
+    # National ID. Compiled case-insensitively: locale packs write letter
+    # classes in upper case (UK NINO, Indian PAN), but IDs are often exported
+    # in lower case, and a missed match classifies the column SAFE, so it is
+    # passed through un-stripped. A miss here is a leak, not a lost warning.
     nid = locale.get("national_id", {})
     nid_type = nid.get("phi_type", "national_id")
     for pat in nid.get("patterns", []):
-        value_patterns.append((re.compile(pat), nid_type))
+        value_patterns.append((re.compile(pat, re.IGNORECASE), nid_type))
 
     # Phone
     for phone in locale.get("phone", []):
@@ -186,9 +189,11 @@ def build_locale_patterns(locale: dict) -> tuple[
         # Build a regex from keywords (case-insensitive word boundary match)
         escaped = [re.escape(kw) for kw in addr_cfg["keywords"]]
         address_re = re.compile(r"(?:" + "|".join(escaped) + r")", re.IGNORECASE)
-    # Postcode pattern (if available, add to value_patterns as address type)
+    # Postcode pattern (if available, add to value_patterns as address type).
+    # Case-insensitive for the same reason as national IDs above.
     if addr_cfg.get("postcode_pattern"):
-        value_patterns.append((re.compile(addr_cfg["postcode_pattern"]), "address"))
+        value_patterns.append(
+            (re.compile(addr_cfg["postcode_pattern"], re.IGNORECASE), "address"))
 
     # Name heuristic
     name_cfg = locale.get("name_heuristic", {})
@@ -382,12 +387,17 @@ def scan_column_names(headers: list[str],
     return results
 
 
-def _sample_values(values: list[str], n: int = 500) -> list[str]:
-    """Return up to n non-empty values for scanning."""
-    non_empty = [v for v in values if v and v.strip()]
-    if len(non_empty) <= n:
-        return non_empty
-    return random.sample(non_empty, n)
+def _values_to_scan(values: list[str]) -> list[str]:
+    """Return every non-empty value in the column.
+
+    This used to be a random sample of at most 500 values, and that caused
+    two defects. First, the same file could classify differently from one
+    run to the next. Second, a column where only a few values carry PHI
+    (three phone numbers in 2,000 notes) came out SAFE whenever the sample
+    missed them, and a SAFE column is passed through un-stripped. A SAFE
+    verdict asserts that no value matched, so every value has to be checked.
+    """
+    return [v for v in values if v and v.strip()]
 
 
 def scan_column_values(col: str, values: list[str],
@@ -406,7 +416,7 @@ def scan_column_values(col: str, values: list[str],
     if name_columns is None:
         name_columns = [k for k, v in UNIVERSAL_COLUMN_NAMES.items() if v == "name"]
 
-    sample = _sample_values(values)
+    sample = _values_to_scan(values)
     if not sample:
         return None
 
@@ -535,7 +545,7 @@ def classify_columns(data: list[dict], headers: list[str],
             if avg_len > 50:
                 # Scan for embedded PHI in free text
                 embedded_phi = False
-                for val in _sample_values(non_empty, 100):
+                for val in non_empty:  # every value; the loop exits on the first hit
                     for regex, _ in val_patterns:
                         if regex.search(val):
                             embedded_phi = True