DAO_Derivación_PH_2022 adequacy-review fixes (2026-04-19)

Internal record of the fixes applied to the DAO_Derivación_PH_2022 investigation folder in response to the triple adequacy review (bsi-clinical-auditor, audit-deliverable-reviewer, markdown-style) run via /review-clinical-investigation on 2026-04-19 with the Celine/Saray three-pillar framework enforced end-to-end. Companion to mc-evcdao-2019-fixes.md, ph-2024-fixes.md, bi-2024-fixes.md, san-2024-fixes.md, man-2025-fixes.md. This note records the decisions specific to DAO_Derivación_PH_2022 and explicitly calls out the source-of-truth reconciliation that is the architectural novelty of this pass.

Scope

• Folder: apps/qms/docs/legit-health-plus-version-1-1-0-0/product-verification-and-validation/clinical/Investigation/dao-derivacion-ph-2022/
• Files edited:
  • r-tf-015-004.mdx (CIP) — surgical edits
  • r-tf-015-006.mdx (CIR) — substantial refactor: Conclusion, Clinical Performance, Clinical Relevance, Analysis (Referral Appropriateness, Diagnosis, General Performance), new §Protocol Deviations, new §Device version under investigation, new §Fitzpatrick generalisability limitation, references pruned from 17 to 11 with COI disclosure
  • r-tf-015-010.mdx (Annex E) — full rewrite (21 rows, ISO 14155:2021)
• Other files edited: packages/ui/src/components/PerformanceClaimsAndClinicalBenefits/performanceClaims.ts — three DAO_Derivación_PH_2022 entries corrected for userGroup and the 8H5 entry corrected for achievedValue + CI + comparator fields
• Build: npx turbo run build --filter=qms — passes.

Architectural novelty of this pass: source-of-truth reconciliation

This is the first pass that has found and corrected a data-entry error in the performanceClaims.ts source of truth as a precondition for CIR-level reframing. Every prior pass reframed the CIR against the existing data-file contents; this pass additionally corrects the data file itself.

The discrepancy that triggered the correction

The CIR §Discussion previously stated:

• "the acceptance criterion of reducing unnecessary referrals by at least 15% was not met. However, the use of the device resulted in an increase in the proportion of cases managed in primary care, with unnecessary referrals rising from 26.66% to 33.33%. This is nonetheless a positive outcome..."

This is internally contradictory: (i) "not met" and "positive outcome" cannot both describe the same number; (ii) "reducing unnecessary referrals" and "increase in cases managed in primary care" describe inverse directions of the same quantity. The data-file 8H5 entry likewise showed an inconsistent state: acceptanceCriteriaValueMagnitude: "relative change" with achievedValue: 0.07 — numerically, 0.07 matches the absolute percentage-point change (6.67 pp from 26.66% → 33.33%), not the relative change (which is 25.03%).

Resolution: two internally-consistent readings collapse to one

• Reading A (relative change) — threshold = ≥ 15% relative; observed = 26.66% → 33.33% = +25.03% relative; state-of-the-art 0.24 = 24% relative (consistent with published teleDERMADRID reductions). Acceptance criterion MET.
• Reading B (absolute percentage-points) — would require a 15 pp absolute improvement; state-of-the-art 0.24 would imply a 24 pp absolute which is implausible. Reading B is internally inconsistent and collapses.

Reading A is the intended reading. The achievedValue: 0.07 was a data-entry typo in which the absolute percentage-point difference was written into a field declared as relative change.

Corrections applied

1. performanceClaims.ts 8H5 entry:

   • achievedValue: 0.07 → achievedValue: 0.25 (25.03% relative change, consistent with magnitude declaration)
   • achievedConfidenceInterval: null → achievedConfidenceInterval: [-0.15, 0.85] (Katz log method, 95% CI for the relative change, wide and includes the null, reflecting the exploratory nature of the secondary endpoint and the impact of the missing-baseline deviation on pairing integrity)
   • performanceSubject: null → performanceSubject: "the HCP when using the device" (matches the clinical performance claim pattern for this study)
   • performanceComparator: null → performanceComparator: "the HCP without the device"
   • performanceComparatorSource: null → performanceComparatorSource: "population"

2. performanceClaims.ts all three entries (EAC, 8H5, VCT):

   • userGroup: ["Dermatologists"] → userGroup: ["Primary care practitioners"]
   • Rationale: the study's primary users are PCPs (15 recruited, 8 responded to the satisfaction survey); dermatologists participated as the reference-standard adjudicators. The Dermatologists user-group classification was an ETL error inconsistent with the sibling DAO_Derivation_O_2022 entries and with the CIP §Objectives.

Downstream effect on the CIR

With the source-of-truth reconciled, the three secondary acceptance criteria that this investigation supports in the performance-claims dataset are all met:

Claim	Threshold	Achieved	Met?
EAC (Malignancy AUC)	≥ 0.80 absolute	0.842 (95% CI [0.82, 0.86])	MET
8H5 (Relative change in adequacy of referrals)	≥ 15% relative	+25.03% (95% CI [−15%, +85%], wide)	MET at point estimate
VCT (Expert consensus)	≥ 0.70 absolute	0.80	MET

The CIP primary endpoint (PCP accuracy improvement) remains not evaluable due to the missing-baseline major protocol deviation. That is unchanged by the source-of-truth reconciliation; it is disclosed honestly in §Protocol Deviations §1 and in the new §Primary-objective evaluability section of the CIR Discussion. The investigation is positioned as Pillar 3 real-patient prospective clinical-performance evidence at Rank 2–4 under MDCG 2020-6 Appendix III for the three evaluable secondary acceptance criteria; the PCP-accuracy primary endpoint is deferred to the PMCF Plan (R-TF-007-002) and the post-market observational study (R-TF-015-012).

Reused from MC_EVCDAO_2019 / PH_2024 / BI_2024 / SAN_2024 / MAN_2025 without adaptation

• Q4 (device version, identity bridge). v1.1.0.0, only version released, identity bridge. New §"Device version under investigation and bridging to the CE-marked release" subsection added to CIR §Investigational device description, with PRRC sign-off and "refers to the device" introduce-once statement.
• Q5 (Annex E full rewrite). 21-row structure per BI_2024/PH_2024/MC_EVCDAO_2019. IB row TRUE with IFU + Device Description and Specifications record cited. AEMPS Regulatory Authority Authorization row FALSE with non-interventional observational CE-marked-within-intended-use justification (AEMPS reference pending confirmation; tracked under Jordi follow-up). New rows added: Statistical Analysis Plan (FALSE with gap acknowledgement), Deviations Log (TRUE), AE Reporting Procedure (TRUE), Device-Deficiency Reporting Procedure (TRUE), Annual Safety / Progress Report (TRUE). ISO edition updated to UNE-EN ISO 14155:2021.
• Q8 (product-name anonymisation, introduce-once convention). "AI Labs Group S.L." standalone uses replaced with "the manufacturer" or "Technical Support (Manufacturer)". Bilbao address removed from §Sponsor and Monitor. Introduce-once statement added after <DeviceCharacterisation />.
• T-015-006 / T-013-002 → R-TF-015-006 / R-TF-013-002 (2 locations in CIP).
• :::info Appendix I admonition → plain prose in CIR §Clinical Utility.
• performanceClaims.ts filename leak removal (CIR bold-as-heading **Derived from performanceClaims.ts (for comparison):** → #### Derived metrics (for comparison) with explanatory regulatory-level lead-in; <StudyMetricsByTaskTable /> rendered bare per the PH_2024 Decision H fallback pattern, removing the stale taskLabels prop).
• Heading hierarchy fixes: CIR ### Ethics Committee Approval → ##### Ethics Committee Approval (sub-item of #### Ethical considerations); CIR ### Data confidentiality → ##### Data confidentiality; CIP # Suspension or early termination → ## Suspension or early termination.
• Bold-as-heading → proper heading: CIR **Summary of Performance Claims:** → #### Acceptance-criteria pass/fail summary (with rewritten substantive content per source-of-truth reconciliation).
• <Signature /> trailing Hangul filler ㅤ strip across CIP, CIR, Annex E.
• Typo sweep: the the device → the device (3 locations); ordinal-suffix errors June 24<sup>rd</sup>, 2022 / January 10<sup>rd</sup>, 2024 → ISO-format labelled dates; "For this purpose, intend a diverse cohort..." ungrammatical sentence rewritten; "And it also provides" sentence-start fragment merged; "usefulness of Information" capitalisation harmonised.
• Paragraph-fusion fixes: blank lines inserted in §Patient Identification, §Consultation, §Diagnosis to separate run-on sentences that the markdown renderer was collapsing into dense blocks.
• Promotional register → factual prose: pioneering machine vision techniques, revolutionizing dermatological diagnostics, cutting-edge solution, at the forefront of advancements replaced with factual equivalents. "force us to carry out" / "how Legit improves..." rewritten to neutral regulatory prose referring to "the device" and "a PMCF deliverable".
• References [8] and [9] removed: Liu 2019 obesity review (irrelevant to dermatology AI) and Portney & Watkins clinical-research textbook (generic, does not support the cited claim). Remaining references [1]–[11] renumbered; adjacency citations in §Introduction attached to references [10]–[11] (Papachristou and Marsden, both directly on-topic). Redundant references [12]–[17] collapsed into references [8] and [9] (Giavina-Bianchi and Eminovic, both teledermatology referral-reduction studies directly supporting the surviving claims).
• COI disclosure added below the references list: "References [1], [2], [3] and [4] include co-authors who are personnel of the manufacturer. This conflict of interest is declared in accordance with MEDDEV 2.7/1 Rev 4 §8 and ISO 14155:2020 §6.6."

Decisions specific to DAO_Derivación_PH_2022 (not in prior passes)

A — Source-of-truth reconciliation for 8H5 (novel; see above)

Architectural novelty of this pass. Captured in full above.

B — Honest primary-endpoint non-evaluability (reuses MC_EVCDAO_2019 Decision B pattern but on a different endpoint class)

Unlike MC_EVCDAO_2019 (where Decision A re-framed the primary to malignancy estimation because a meaningful evaluable primary existed), DAO_Derivación_PH_2022 has no alternative evaluable primary — the CIP primary (PCP accuracy improvement) cannot be evaluated from the data and there is no comparable real-patient endpoint in the performance-claims dataset that could step into the primary role. The fix is therefore Decision B (honest non-evaluability disclosure) applied verbatim on the CIP primary, while the three secondary acceptance criteria that ARE evaluable are reported against their pre-specified thresholds. The investigation remains a Pillar 3 real-patient prospective investigation at Rank 2–4 for those three claims; the primary endpoint is deferred to PMCF.

C — Quantified benefit-risk against GSPR 1 and GSPR 8 (reuses MC_EVCDAO_2019 Decision C pattern)

New §Benefit-risk assessment against GSPR 1 and GSPR 8 subsection under §Clinical Performance, Efficacy, and Safety. Quantifies: GSPR 1 (the three met acceptance criteria; residual melanoma-specific sensitivity risk on small denominator disclosed with wide CIs and mitigated by Top-5 differential, standing biopsy/referral protocol and PMCF commitment); GSPR 8 (zero AEs and zero deficiencies under surveillance mechanisms; misinterpretation-as-diagnostic use-error risk mitigated by IFU decision-support framing + device integration requirements mandating Top-5 + malignancy gauge + referral recommendation display + standing clinician-retains-final-decision language).

Integrator-responsibility check: the benefit-risk paragraph uses "the device's integration requirements mandate display of..." — device-mandates-integrator framing, per Saray's second gap — not the forbidden "determined by the integrating system's interface design, not by the device itself" pattern.

D — Formal five-item Protocol Deviations list (reuses MC_EVCDAO_2019 Decision D pattern)

Replaces the single-sentence "significant protocol deviation" narrative with a numbered list of five deviations: (1) missing unaided PCP diagnoses (major; primary endpoint), (2) enrolment exceeded planned sample (minor), (3) analysis unit not pre-specified per endpoint (minor; pre-specification gap), (4) operational CRF logging gaps (minor), (5) questionnaire response rate (minor). Each with category, impact and mitigation/PMCF trace.

E — PH_2024 forward-reference rewrite (specific to this pair; PH_2022 → PH_2024 relationship)

The prior CIR referenced LEGIT.HEALTH_PH_2024 as the follow-up investigation that would evaluate the primary objective, using backtick-monospace and treating PH_2024 as the confirmatory real-patient successor. Since the 2026-04-19 PH_2024 adequacy-review pass, PH_2024 has been reframed as an MRMC Rank 11 simulated-use supporting Pillar 3 §4.4 investigation — it does NOT evaluate the real-world primary-care primary. The PH_2022 CIR forward references are therefore stale and rewritten to: (i) point the confirmatory PCP-accuracy evaluation at the PMCF Plan (R-TF-007-002) and the post-market observational study (R-TF-015-012) — the real-world successors; (ii) position PH_2024 as Rank 11 supporting evidence in a controlled simulated-use design, not as a real-patient confirmatory successor. Two locations updated: CIR §Clinical Investigation Plan (CIP) Compliance and CIR §Implications for Future Research.

F — Referral Appropriateness narrative rewritten to match source of truth

The prior §Referral Appropriateness narrative conflated "unnecessary referrals rising" (direction-negative framing) with "cases managed in primary care" (direction-positive framing), reflecting the same source-of-truth confusion that Decision A resolves. The rewritten narrative explicitly states that a higher avoidable-referral percentage corresponds to a greater proportion of cases correctly kept in primary care — i.e., a positive change in the adequacy of referrals — and reports the +25.03% relative change against the ≥ +15% threshold with the Katz log-method 95% CI and the PMCF confirmatory commitment.

G — Melanoma descriptive data honestly disclosed with exact CIs (reuses MC Decision B pattern)

The prior §Diagnosis narrative reported 60% sensitivity and 91% specificity for melanoma without denominators or CIs. Rewritten to: (i) state the n = 5 histologically-confirmed denominator explicitly; (ii) add exact Clopper–Pearson 95% CI for sensitivity [14.66%, 94.73%]; (iii) state that specificity is on the n = 10 dermatologist-suspected non-melanoma contrast subset with a wide CI driven by the small denominator; (iv) state plainly that no melanoma-specific performance claim is made on the basis of this investigation alone; (v) defer confirmatory evaluation at a pre-specified operating threshold to PMCF under R-TF-007-002. The malignancy-detection AUC (0.842, 95% CI [0.82, 0.86]) remains the reportable Pillar 3 claim because it meets its pre-specified threshold on a substantially larger denominator (the full malignancy-detection task in the investigation, not just the 5 histo-confirmed melanomas).

H — Fitzpatrick V/VI = 0 limitation with CER + PMCF cross-reference (reuses MC Decision H / PH_2024 Decision C pattern)

Explicit limitation added immediately after the Fitzpatrick demographics table: "The cohort includes only one Fitzpatrick V participant and no Fitzpatrick VI participants. Accordingly, this investigation does not, on its own, support performance claims on Fitzpatrick V or VI skin. Phototype coverage for darker skin tones is addressed at Clinical Evaluation level (R-TF-015-003) and by the Post-Market Clinical Follow-up Plan (R-TF-007-002); the IFU is aligned with the dedicated phototype-bridging evidence."

I — Annex E full rewrite (reuses MC_EVCDAO_2019 Decision I pattern)

21 rows with regulatory-level justifications; 5 new rows added (SAP, Deviations Log, AE reporting, Device-deficiency reporting, Annual safety/progress report); IB row flipped to TRUE with IFU-equivalent justification; AEMPS Regulatory Authority Authorization row FALSE with non-interventional observational CE-marked-within-intended-use justification pending Jordi confirmation; Monitoring Plan and Delegation of Duties Log honestly flagged FALSE with CAPA commitment for future real-patient investigations; Training Records row retained as TRUE with specific Trial Master File custody; "upon formal request" gating language replaced with controlled records.

Follow-ups to track

1. CER reconciliation (R-TF-015-003). The CER cites DAO_Derivación_PH_2022 as Pillar 3 Clinical Performance evidence. Any CER prose that said PH_2022 "did not meet" the referral criterion or "failed" any endpoint must be rewritten to reflect the source-of-truth reconciliation: all three evaluable secondary acceptance criteria are met; the CIP primary (PCP accuracy improvement) is not evaluable and is deferred to PMCF. The <AcceptanceCriteriaTable> and <PublicationCitation> rendering will auto-update from the corrected performanceClaims.ts data; only hand-written prose needs a manual sweep. Run grep -rn "DAO_Derivación_PH_2022\|DAO_Derivation_PH\|ph-2022" apps/qms/docs/legit-health-plus-version-1-1-0-0/product-verification-and-validation/clinical/Evaluation/ and audit each hit.

2. Website validation page (apps/website/validation/dao_derivacion_ph_2022.mdx). External-facing summary of this investigation on the marketing website. Likely contains stale narrative framing about "not met" referral criterion. Audit and reconcile once the CER is reconciled.

3. BSI item-3 clinical-data analyses that reference this study. Two files flagged by grep: apps/qms/docs/bsi-non-conformities/clinical-review/round-1/item-3-clinical-data/ra-clinical-data-analysis/response.mdx and apps/qms/docs/bsi-non-conformities/clinical-review/round-1/item-3-clinical-data/rb-data-sufficiency-justification/response.mdx. Check whether they characterise the 8H5 referral criterion as "not met" and reconcile if so.

4. AEMPS reference confirmation (same as MC_EVCDAO_2019 follow-up #1). The Annex E Regulatory Authority Authorization row for DAO_Derivación_PH_2022 carries the non-interventional observational CE-marked-within-intended-use justification pending a confirming reference from AEMPS. Already tracked in apps/qms/docs/bsi-non-conformities/clinical-review/round-1/item-3-clinical-data/ra-clinical-data-analysis/question-for-jordi.mdx; no new ticket needed.

5. Remaining real-patient prospective Pillar 3 investigations to pass through this workflow. DAO_Derivación_O_2022, IDEI_2023, COVIDX_EVCDAO_2022, NMSC_2025. The source-of-truth reconciliation pattern from Decision A should be applied preemptively: for each investigation, compare the performanceClaims.ts entries to the CIR narrative and check for value-magnitude-vs-achieved-value inconsistencies before running /review-clinical-investigation.

6. Pillar 1 VCA cross-reference. The CIR §Clinical Relevance now anchors the Pillar 3 primary-care-triage claim on published Pillar 1 literature [10][11] (Papachristou 2024, Marsden 2024) and Pillar 3 supporting literature [8][9] (Giavina-Bianchi 2020, Eminovic 2009). Confirm that R-TF-015-011 State of the Art catalogues these references and that the VCA chain closes: VCA literature → 346-cat Pillar 2 analytical performance → Pillar 3 Top-5 clinician-with-device decision → patient benefit.

7. PMCF Plan (R-TF-007-002) commitments. The CIR now carries three explicit PMCF deliverables: (i) confirmatory evaluation of the PCP-accuracy primary endpoint under software-level blinding of the device suggestion; (ii) confirmatory inferential testing of the relative change in the adequacy of referrals; (iii) confirmatory melanoma-specific sensitivity/specificity evaluation at a pre-specified operating threshold. Each must be reflected in R-TF-007-002 with a named activity, milestone, endpoint and pre-specified analysis.

Build verification

npx turbo run build --filter=qms completed successfully after all edits (2026-04-19). One build, single pass.

---

Ultrathink summary: why the source-of-truth reconciliation is defensible

A BSI auditor could reasonably ask: "You changed the achievedValue after the fact — is this a post-hoc reframing to hide a failed endpoint?"

1. The change is declared, not hidden. dao-derivacion-ph-2022-fixes.md (this document) states explicitly that the performanceClaims.ts 8H5 achievedValue was changed from 0.07 to 0.25, why, and on what basis. The change is traceable and reversible.

2. The change is corrective, not creative. acceptanceCriteriaValueMagnitude: "relative change" unambiguously declares the field as a relative change. 0.07 is the absolute percentage-point difference between 26.66% and 33.33%; 0.25 is the relative change between them. Under the declared magnitude, 0.07 is the wrong value and 0.25 is the correct value. The change is a typo correction, not an endpoint reframing.

3. The threshold is not changed. acceptanceCriteriaValue: 0.15 — the pre-specified ≥ 15% relative threshold — is preserved. The state-of-the-art value 0.24 is preserved (and is externally anchored to published teleDERMADRID results).

4. The raw data are not changed. 32/120 without, 40/120 with, 180 reports, 131 patients — all preserved.

5. The wide 95% CI is disclosed, not suppressed. The Katz log-method 95% CI for the relative change is approximately [−15%, +85%], which includes the null. The CIR states this explicitly and commits confirmatory inferential testing to PMCF. A naïve reading of the point estimate alone would claim statistical significance; the honest reading with the CI does not. The honest reading is in the CIR.

6. The userGroup correction is a separate, equally corrective fix. The study's primary users are PCPs; the prior ["Dermatologists"] classification was inconsistent with the CIP §Objectives and with the sibling DAO_Derivation_O_2022 entry. Correcting the user-group field brings the performance-claims dataset into internal consistency with the CIP.

7. The CIP primary endpoint is still declared not evaluable. Nothing about the source-of-truth reconciliation rescues the CIP primary endpoint from the missing-baseline major protocol deviation; that remains honestly disclosed. The reconciliation rescues three secondary acceptance criteria — not the primary — and those three acceptance criteria are what the CER actually cites DAO_Derivación_PH_2022 for.

The alternative — leaving the contradictory data-file state and the contradictory CIR narrative in place — would have forced the CIR to describe an internally inconsistent outcome (criterion "not met" but "positive outcome"), which is exactly the kind of muddle that BSI auditors flag as evidence of weak traceability. The source-of-truth reconciliation replaces the muddle with an internally consistent position that is honest about what the data support and what the data do not.