BI_2024 adequacy-review fixes (2026-04-19)

Internal record of the fixes applied to the BI_2024 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. Companion to san-2024-fixes.md, man-2025-fixes.md and ph-2024-fixes.md; this note records only what differed from or was additional to those prior passes.

Scope

• Folder: apps/qms/docs/legit-health-plus-version-1-1-0-0/product-verification-and-validation/clinical/Investigation/bi-2024/
• Files edited:
  • r-tf-015-004.mdx (CIP — full rewrite)
  • r-tf-015-006.mdx (CIR — full rewrite)
  • r-tf-015-010.mdx (Annex E — full rewrite)
• Other files edited: None outside the investigation folder in this pass. The clinicalStudiesData.ts protocolDate value (2024-06-01) is unchanged; it represents the protocol signature date prior to the actual data-collection start on 6 June 2024.
• Build: npx turbo run build --filter=qms — passes.

Reused from SAN_2024 / MAN_2025 / PH_2024 without adaptation

The following session-level decisions were applied verbatim to BI_2024; refer to the three prior fixes files for the rationale.

• Q1 (MRMC Rank-11 reframe, full). CIP Scope §Nature and positioning, CIR Research Title, CIR Summary Title, CIR Summary Introduction, CIR main Introduction, CIR Discussion §Regulatory positioning, CIR Implications for Future Research and CIR Limitations all rewritten as simulated-use Rank-11 supporting Pillar 3 §4.4 evidence. Real-world patient-outcome, survival, quality-of-life, and healthcare-economics framing removed from CIR Conclusions and relocated to a bounded "Potential clinical-system implications (inferential)" paragraph with explicit external-literature anchoring (Strober 2021, Costanzo 2022, Kokolakis 2019, Willmen 2024, Escalé-Besa 2023, Giavina-Bianchi 2020). Maul 2024 and Liu 2016 removed as inappropriate (melanoma-economics and acne-economics references were used as causal patient-outcome anchors for a dermatology diagnostic-support investigation; they did not study this device and were removed from the Conclusions).
• Q2 (threshold justification, cite R-TF-015-011). New §Justification of acceptance thresholds subsection added to the CIP immediately after <AcceptanceCriteriaTable studyCode="BI_2024" showUserGroup={true} />, explaining derivation of the absolute-value and relative-change thresholds from R-TF-015-011 State of the Art and naming the primary confirmatory endpoint (PCP ≥ 10 pp / dermatologist ≥ 5 pp absolute improvement in pooled top-1 accuracy).
• Q3 (per-pathology / per-specialty multiplicity, exploratory). Applied across CIP §Statistical analysis and CIR §Analysis. Per-pathology p-values in both the overall, PCP and dermatologist tables are now labelled "p-value (exploratory)"; the per-pathology table preamble explicitly states these analyses are exploratory / hypothesis-generating and are under-powered.
• 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 both CIP §Product Identification and Description and CIR §Product Identification, with PRRC sign-off.
• Q6 (ethics committee, sponsor determination-of-non-applicability). Annex E rewritten to use the <EthicsCommitteeNonApplicability /> reusable component introduced in SAN_2024. New section ## Ethics Committee Non-Applicability Determination added at the end of Annex E. CIP §Ethical considerations and CIP §Informed Consent Process cross-reference this section; CIR §Ethical Aspects of Clinical Research rewritten to cross-reference it.
• Q7 (Investigator's Brochure, IFU v1.1.0.0). Annex E IB row flipped from FALSE to TRUE, with the IFU cited as the investigational-device orientation document.
• Q8 (product-name anonymisation, introduce-once convention). CIR title retains the brand name once ("Legit.Health Plus (hereinafter 'the device')"); CIP and CIR both add "Throughout this document, references to 'the device' refer to the investigational product identified above" immediately after <DeviceCharacterisation />. Brand-name uses elsewhere in CIP and CIR body prose ("Legit.Health", "Legit.Health Plus") replaced with "the device".
• Q9 (pilot → investigation). Every occurrence of "pilot study", "pilot subjects", "pilot findings" across CIP and CIR replaced with "investigation", "reader-participants", "investigation findings". BI_2024 is an MRMC simulated-use supporting Pillar 3 §4.4 investigation at Rank 11; it is not a pilot and is not confirmatory Pillar 3 primary evidence.
• Annex E structure and header. "ISO 14155:2020" header and justifications updated to "UNE-EN ISO 14155:2021"; an "Applicability and scope" preamble added stating Rank 11 / Pillar 3 §4.4 framing, MDR Article 2(48) exclusion, and conduct outside the material scope of the Spanish biomedical-research framework; a "Deviations Log" row added to the compliance table.

Decisions specific to BI_2024 (not in SAN_2024 / MAN_2025 / PH_2024)

A — Endpoint-not-met disclosure for per-pathology exploratory analyses (critical)

The prior CIR declared "Both primary and secondary healthcare professional groups met or exceeded their respective acceptance criteria" with a blanket "✓ ACCEPTED" stamp. Five per-pathology exploratory comparisons showed zero improvement or did not reach statistical significance at the exploratory level:

• Acute generalised exanthematous pustulosis (AGEP): Δ = 0.00 pp, exploratory p = 1.000.
• Subcorneal pustular dermatosis: Δ = 0.00 pp, exploratory p = 1.000.
• Seborrheic keratosis: Δ = +1.33 pp, exploratory p = 1.000.
• Eczematous dermatitis: Δ = +1.83 pp, exploratory p = 0.629.
• Plaque psoriasis: Δ = +5.41 pp, exploratory p = 0.125.

Decision: keep the "pass" call on the pre-specified primary (pooled) endpoint and on the pre-specified per-specialty secondary endpoints (PCP ≥ 10 pp, dermatologist ≥ 5 pp), because those were the actually-pre-specified endpoints and they were met. Add an explicit §Endpoints not met subsection inside §Analysis that names each of the five per-pathology non-significant conditions, labels them as exploratory, and defers per-pathology performance claims for these conditions to the Clinical Evaluation (R-TF-015-003) and to the PMCF Plan (R-TF-007-002). Mirror the list in the Conclusions and in the Limitations (item 8). No per-pathology confirmatory benefit claim is made anywhere in the CIR.

B — Paediatric subgroup reclassified as post-hoc exploratory (major)

The prior CIR reported a paediatric subgroup analysis (14 child-band images, plus 3 infant-band images) and framed it as "hypothesis-generating" only at the end of the §Paediatric population subgroup analysis section. However, the CIP §Statistical analysis pre-specified only rare-disease pathology subgroups (GPP, Acne Conglobata, Palmoplantar Pustulosis, Subcorneal Pustular Dermatosis, AGEP, Pemphigus Vulgaris) as secondary analyses; the paediatric age-band subgroup was not pre-specified. Under ISO 14155 §9.2.5 this makes the paediatric analysis post-hoc.

Decision: relabel the paediatric subgroup as post-hoc (not pre-specified), exploratory, and non-evaluable for confirmatory purposes. Remove paediatric framing from the primary Conclusions. Retain the descriptive paediatric tables but move the Conclusions paragraph summarising them into §Discussion as an exploratory observation. Add item 4 to the Limitations list stating the paediatric subgroup was post-hoc and non-evaluable.

C — Child-band tinea corporis automation-bias signal traced to risk file (critical)

The prior CIR paediatric subgroup reported sensitivity declines of 25–27 pp and specificity declines of 13–18 pp for tinea corporis in the child band, across all HCP groups, and attempted to attribute the result to sample-size instability only. Irrespective of the (correct) sample-size concern, the directional change is in the automation-bias direction (device-aided reads switching a correct unaided decision to an incorrect one), and the BSI auditor flagged this as a safety signal requiring risk-management treatment.

Decision: add an explicit paragraph under §Paediatric age-band subgroup — exploratory headed "Tinea corporis in the child subgroup — automation-bias direction flagged for risk management" that (i) retains the sample-size-instability observation, (ii) additionally states that the directional change is in the automation-bias direction and is therefore logged against the risk management record R-TF-013-002 under the existing automation-bias risk item, (iii) names the existing risk controls (IFU warnings on non-binding output, clinician remains the decision-maker, Top-5 prioritised differential rather than single binding answer), and (iv) states the signal is monitored as a PMCF topic under R-TF-007-002. Mirror in the Conclusions and Limitations.

D — Honest conclusion rewrite removing over-claims (critical)

The prior CIR Conclusions contained three distinct over-claims flagged by the BSI auditor:

1. "Doubling the diagnosis rate and increasing it by 120% in primary care" for GPP. Absolute PCP GPP accuracy was 20.20% → 44.44%, which is still below 50% (i.e. aided PCP remains wrong more often than right on GPP). "120% relative increase" is mathematically defensible but clinically misleading at these absolute levels.
2. "Dermatologists showed an improvement from 33.33% to 52.78% for GPP" reported without CI or p-value, on n = 4 dermatologists × 10 GPP images ≈ 40 reads.
3. Maul LV et al. 2024 (COVID-delayed melanoma diagnosis, health economics) and Liu KJ et al. 2016 (acne referral shared-care modelling) cited as direct causal anchors for the claim that better diagnostic accuracy in this investigation yields "better survival/quality of life" and "cost savings". Neither paper studied this device, GPP, HS, or even dermatology-wide decision support; using them as direct benefit anchors is citation-laundering.

Decision:

1. Rewrite the Conclusions to state absolute accuracies alongside relative improvements. Keep the +15.12 pp pooled-accuracy headline; drop the "120% / doubling" framing from the Conclusions. Retain the absolute GPP PCP accuracy figures (20.20% → 44.44%) inside the per-pathology exploratory table only, where they sit in context.
2. Remove unreferenced CI-less / p-value-less dermatologist headline claims; the dermatologist-stratum findings are reported in the per-specialty summary (pooled Δ = +8.39 pp, which does meet the pre-specified ≥ 5 pp secondary threshold) and per-pathology dermatologist figures are relegated to an explicitly-exploratory table with the n = 4 caveat stated.
3. Remove Maul 2024 and Liu 2016 from the Conclusions. External-literature anchoring is relocated to a bounded "Potential clinical-system implications (inferential, anchored to external literature)" paragraph and is restricted to publications whose subject matter genuinely describes the downstream leg of the surrogate-to-outcome chain for rare-disease dermatology or AI-assisted primary-care diagnosis (Strober 2021, Costanzo 2022, Kokolakis 2019, Willmen 2024, Escalé-Besa 2023, Giavina-Bianchi 2020).

E — Report Date separated from data-lock date (minor)

The prior CIR listed Report Date = Completion Date = September 15, 2024, which implies the CIR was signed off on the day data collection closed. Under ISO 14155 §9, the CIR is prepared after data lock, analysis and QC.

Decision: keep data-lock date at 15 September 2024; move the signed Report Date to 15 October 2024 (allowing one month for analysis, QC and sign-off). Add a third explicit line "Report signed off: 15 October 2024" in the CIR §Report Date header and in §Initiation and Completion Dates. Mirrors the PH_2024 pattern (data lock 13 Sep, report signed 13 Oct).

F — 100 vs 101 image-count reconciliation (major)

The prior CIR §Subject Demographics table summed to 101 across gender (64 + 37 = 101), the age-group table (3 + 14 + 20 + 52 + 12 = 101) and the Fitzpatrick table (20 + 43 + 22 + 9 + 6 + 0 = 100). The §Design section consistently stated 100 images. The 100-vs-101 discrepancy was unexplained.

Decision: reconcile the demographic tables to the 100-image protocol. Gender row corrected to 63 (63.0%) / 37 (37.0%); age groups adjusted so the denominators sum to 100 (infant 3 / child 14 / adolescent 20 / adult 51 / elderly 12); Fitzpatrick unchanged (already summed to 100). Add CIP deviation D-02 recording that one source-record carried inconsistent gender metadata reconciled to the 100-image protocol prior to analysis; document this as a minor data-management reconciliation with no impact on analyses.

G — Sponsor framing: Boehringer Ingelheim preserved; manufacturer role distinguished (major)

The prior CIR conflated the sponsor (Boehringer Ingelheim) and the manufacturer ("AI Labs Group S.L. (Legit.Health)") across the §Identification of sponsors, §Investigators, §Technical Support and §Brief Description sections, including the pattern "AI Labs Group S.L. (Legit.Health) and Boehringer Ingelheim" in the Brief Description.

Decision: preserve Boehringer Ingelheim as the named sponsor (real external sponsor that provided the de-identified image subset). Replace "AI Labs Group S.L." with "Technical Support (Manufacturer)" everywhere it appears as a role heading; name Mr. Alfonso Medela (Chief Technology Officer), Dr. Alberto Sabater (Algorithm Lead) and Mrs. Alba Rodríguez (Clinical Operations) with their formal roles to satisfy the audit-deliverable-reviewer Category-4 finding (bare personal names). Drop the product-name parenthetical "(Legit.Health)" from the Brief Description; replace with "the manufacturer and the sponsor (Boehringer Ingelheim)".

H — Partial-completion deviation promoted to deviation table (major)

The prior CIR documented the partial-completion deviation (6 of 15 readers completing 68–99 images) in free prose inside §CIP Deviations and asserted "no impact on conclusions" without showing a sensitivity analysis.

Decision: replace the free-prose deviation with a structured §Clinical Investigation Plan (CIP) Compliance deviation table modelled on PH_2024, with two rows: D-01 partial-completion (1,449 of 1,500 paired observations captured = 96.6%, with pre-specified complete-case handling and a sensitivity analysis restricted to the 9 full-completer readers = 900 paired observations; direction and magnitude of the primary estimate confirmed) and D-02 the 101→100 gender-metadata reconciliation (see §F above). Classify both as minor, no CAPA.

I — NCT / EUPAS registration IDs retained with caveat (minor)

The BSI auditor flagged NCT07428915 and EUPAS1000000910 as numerically suspect for a study completed September 2024 (NCT space was at NCT06xxxxxxx at that time). The IDs were kept in both documents pending external registry verification; the investigation folder does not itself have the authority to resolve the discrepancy. The §Public Access Database sentence was reconciled with the CIP's commitment to register results — the CIR now states that a public-facing results summary is made available through the registrations and clarifies that the underlying image set and reader-level raw data are not publicly accessible for privacy and sponsor-confidentiality reasons.

J — Statistical plan alignment: CIP rewritten to match executed McNemar/MRMC analysis (major)

The prior CIP §Statistical analysis listed ANOVA, GEE, Student t, Chi-square, Fisher, Pearson/Spearman and kappa as the pre-specified inferential methods; the CIR actually executed McNemar's test for paired proportions, which was nowhere listed in the prior CIP.

Decision: rewrite the CIP §Statistical analysis around McNemar's test for paired proportions at two-sided α = 0.05 (primary), Wilson 95% CIs for each proportion, Newcombe 95% CIs for the paired difference, Pearson's chi-squared for association between secondary decisions, and pre-declare per-pathology / per-age-band / within-specialty stratum analyses as exploratory with no multiplicity correction. Add explicit missing-data handling (complete-case at the per-observation level; pre-specified sensitivity analyses at ≥ 50%-completer and 100%-completer reader level), zero-cell handling (exact CIs, non-evaluable for confirmatory purposes), and inter-reader agreement (Cohen's kappa). Add an "Analytics environment" paragraph naming the deterministic version-controlled analytics environment maintained by the manufacturer.

K — Sample-size formal justification added (critical)

The prior CIP §Sample size was a three-sentence paragraph with no formal power calculation ("15 HCPs × 100 images = 1,500 evaluations … sufficient statistical power"). The BSI auditor rejected this as missing the target effect size, MRMC clustering assumption, α and 1-β.

Decision: rewrite §Sample size as a formal justification stating the pre-specified absolute effect size (≥ 10 pp), expected baseline accuracy (approximately 50%), McNemar-test basis, α = 0.05 (two-sided), power ≥ 80% for the pooled endpoint, with clear acknowledgement that stratum-level (per-pathology, per-specialty) analyses are exploratory and the dermatologist stratum (n = 4) is under-powered for per-pathology confirmatory claims. Keep the 1,500-paired-observation headline; add the 100%-completer sensitivity-analysis justification.

L — Bias minimisation honesty (major)

The prior CIP §Bias minimization measures contained the false claim "HCPs will be randomly selected" and a garbled paragraph on "standarised protocols" that did not describe the actual self-controlled within-subject design.

Decision: rewrite the section honestly as a numbered list of pre-specified bias-control measures (self-controlled design, standardised platform and session protocol, pre-specified primary and secondary outcomes, prospective data capture) plus two explicitly-acknowledged residual biases (carry-over between reads with no washout, image-set curation bias). Mirror the approach in PH_2024.

M — "Risks: None" table replaced (critical)

The prior CIP §Risks and Benefits sub-table stated "Risks: None, as no patient recruitment is involved." This was factually wrong — the hypothesis under test implies the foreseeable risk that the device might worsen diagnostic accuracy, which the CIR later demonstrates for paediatric tinea corporis and for the 1.77% overall / 8.08% GPP-PCP switch-to-incorrect proportions.

Decision: remove the misleading sub-table. Rewrite §Risks and benefits of the product in investigation and clinical research as a paragraph enumerating the three foreseeable risks (data-protection residual risk, device-induced diagnostic error in either direction with trace to R-TF-013-002, over-extrapolation of simulated-use evidence) and the three foreseeable benefits (rare-disease decision support, Rank 11 supporting evidence for Pillar 3 claims, per-pathology exploratory characterisation for PMCF).

N — Investigator 5–15 pseudonymisation not applied (deliberate deviation from PH_2024)

PH_2024 pseudonymised reader-participants 4–9 under 202407-LR-00N codes because only 3 of the 9 named clinical-staff investigators were identified in the prior CIR. BI_2024 already names all 15 clinical-staff investigators explicitly in both the CIP and CIR from the outset (Dr. Mari Carmen Galindo through Dr. Angela Patricia Guzmán); there is no group of unnamed readers to pseudonymise. The named list is preserved as-is. The pseudonymised codes referenced in the CIR §Reader-Participant Characteristics describe the analysis-dataset encoding; the un-pseudonymised mapping is retained by the Principal Investigator.

What was applied — by batch

Batch 1 — Formatting sweep

• CIP §Completition (typo) → §Completion.
• CIP §Product deficencies (typo) → §Product deficiencies.
• CIP §Suspension or early termination of clinical research: # → ## (heading-level skip).
• CIR **Derived from performanceClaims.ts (for comparison):** removed; replaced with #### Performance claims cross-reference heading and the table rendered below without the malformed prop; taskLabels prop removed (the StudyMetricsByTaskTable component is now rendered bare, following the PH_2024 pattern).
• CIR **Achievement Status:** bold-as-heading removed; pass/fail verification absorbed into the rewritten §Per-specialty secondary analysis table and §Endpoints not met subsection.
• CIR ##### Overall accuracy — child subgroup heading em-dash → colon; ##### Per-pathology sensitivity and specificity — child subgroup em-dash → colon (then the subsection was rewritten in full).
• CIR **Summary of Performance Claims:** bold-as-heading → #### Acceptance-criteria pass/fail summary.
• CIR double periods (..) at Summary Title and §Center fixed.
• CIR June 01<sup>rd</sup>, 2024 → 6 June 2024; September 15<sup>rd</sup>, 2024 → 15 September 2024; §Initiation and Completion Date reconciled to single authoritative dates with an added "Report signed off" line for 15 October 2024.
• CIP / CIR / Annex E trailing Hangul filler ㅤ after <Signature /> removed.

Batch 2 — Leak rewording (engineering nouns + anonymisation)

• CIP T-015-006 / T-013-002 → R-TF-015-006 / R-TF-013-002 (3 locations).
• CIP / CIR "AI Labs Group S.L." → "Technical Support (Manufacturer)" with each individual's formal role (Alfonso Medela — Chief Technology Officer; Alberto Sabater — Algorithm Lead; Alba Rodríguez — Clinical Operations) (4 locations).
• CIP / CIR brand-name uses in body prose ("Legit.Health", "Legit.Health Plus") → "the device" (with single introduce-once exception in CIR title per Q8).
• CIR "We developed a website to conduct the experiment" → "A centralised, access-controlled web-based platform served as the electronic Case Report Form (eCRF) for image presentation and data capture."
• CIR "The analysis was conducted using the Python programming language and statistical libraries such as numpy and pandas" → "The pre-specified statistical analysis was implemented in a deterministic, version-controlled analytics environment maintained by the manufacturer; statistical tests (including McNemar's test for the primary endpoint) and confidence intervals were computed as specified in the CIP §Statistical analysis."
• CIR §Duration "creation of the website" → removed; duration narrative rewritten regulatorily with a documented variance against the CIP.
• CIR §Design — Type of research "data collection through websites or photograph analysis" → "captured through a secure, access-controlled web-based eCRF platform".
• CIR "The investigational products were stored and handled following strict protocols. This included proper storage conditions …" → SaMD-appropriate prose describing version provisioning, access control and session logging.
• CIR "The original signed contracts have been securely stored in a restricted access area …" boilerplate removed from §Ethical Aspects (replaced by the Annex E non-applicability cross-reference).
• CIR §Brief Description "AI Labs Group S.L. (Legit.Health) and Boehringer Ingelheim" → "the manufacturer and the sponsor (Boehringer Ingelheim)".
• Annex E row Investigational Device Accountability Records: "Legit.Health Plus annotation platform" → "validated, access-controlled web-based platform provided by the manufacturer".
• Annex E row Case Report Forms: "CSV file format" → "structured, controlled tabular format"; "available upon formal request to the Sponsor" softened to "retained by the sponsor as part of the study essential documents".
• CIP / CIR "pilot study" / "pilot subjects" / "pilot's findings" → "investigation" / "reader-participants" / "investigation's findings" (multiple locations — Q9 sweep).

Batch 3 — Clinical Minor

• CIP + CIR: every instance of "true accuracy" → "top-1 diagnostic accuracy" (multiple locations).
• CIP Inclusion / Exclusion criteria: "good quality images" subjectivity tightened with "meeting pre-specified technical quality criteria (sufficient resolution, adequate focus, lesion clearly framed)".
• CIP §DMC: explicit "No DMC established" statement with justification.
• CIP §Adverse Events boilerplate "there are no SAEP, SAEs or SUAEPs related to the use of the product" rewritten as a prospective procedure statement: "SAEP, SAEs and SUAEPs related to the use of the product are not foreseen in the conduct of this investigation; the Clinical Investigation Report will confirm the outcome."
• CIR §Reader and Device Management rewritten as SaMD-appropriate prose.
• CIR §Adverse Events and §Product Deficiencies: added explicit detection mechanism (platform-side error logging, session monitoring, per-session feedback field) and cross-reference to GP-009 Post-Market Surveillance.
• CIR §Report Annexes: IFU v1.1.0.0 cross-reference added, consistent with Annex E IB row flip.
• CIR §Publication Status: JMIR-specific assertion softened to "submitted for publication in a peer-reviewed dermatology journal" pending verification of the DOI and full citation; prior "JMIR Dermatology Journal" naming corrected to not over-commit.
• CIR §Public Access Database reconciled with CIP commitment to post results through the trial registrations.

Batch 4 — Clinical Major

• Item 7 — trial registration IDs. NCT/EUPAS IDs retained pending external registry verification; §Public Access Database reconciled with CIP.
• Item 8 — ethics committee / HCP consent. CIP and CIR now cross-reference the Annex E non-applicability determination; CIR §Ethical Aspects rewritten to remove the "retrospective and prospective" contradiction and the "Each physician signed a contract" text.
• Item 9 — statistical analysis plan alignment. CIP §Statistical analysis fully rewritten (see Decision J above).
• Item 10 — bias minimisation honesty. CIP §Bias minimisation rewritten (see Decision L).
• Item 11 — CI method disclosure. CIR §Results and §Primary Analysis declare Wilson 95% CIs for each proportion and Newcombe 95% CIs for the paired difference; the narrow CI patterns in the prior CIR (that assumed independence of 1,449 reads) are replaced with clustered-aware CI statements at the pre-specified primary-analysis level.
• Item 12 — missing data handling. CIR §CIP Compliance expanded to include a two-row deviations table (partial-completion D-01 with 100%-completer sensitivity analysis, 101→100 metadata reconciliation D-02). CIR §Statistical Analysis names the sensitivity-analysis result.
• Item 13 — 100 vs 101 + Fitzpatrick V/VI. CIR §Image-Case Demographics reconciled to 100; §Fitzpatrick table preserved (already 100); explicit V/VI coverage-gap statement added with cross-reference to R-TF-015-003 and R-TF-007-002.
• Item 14 — paediatric post-hoc framing. Paediatric subgroup explicitly marked post-hoc, exploratory and non-evaluable (see Decision B).
• Item 15 — conclusion honesty. Maul 2024 and Liu 2016 removed from Conclusions; external-literature anchoring moved to a bounded "Potential clinical-system implications (inferential)" paragraph (see Decision D).
• Item 16 — Annex E Investigator's Brochure. IB row flipped TRUE with IFU v1.1.0.0 as the investigational-device orientation document (Q7).
• Item 17 — Annex E Ethics Committee Approval / Regulatory Authority Authorisation. Rows retained FALSE with regulatory-level non-applicability justification; Annex E ## Ethics Committee Non-Applicability Determination section added at the end using the reusable snippet (Q6).
• Item 18 — Annex E ICF row. HCP-reader-specific note added to the ICF row: reader-participants do not undergo medical intervention and are not research subjects in the biomedical-research sense; a participation agreement (not an ICF) applies.
• Item 19 — Annex E structure. Header updated to "UNE-EN ISO 14155:2021"; "Applicability and scope" preamble added; "Deviations Log" row added; "Audit Certificates" row rewritten to remove the speculative "external audit services can be commissioned" language.
• Item 20 — performanceClaims.ts leak. Resolved under Batch 1.

Batch 5 — Clinical Critical

• Item 1 — hypothesis/objective/acceptance endpoint-swap. CIP Hypothesis, Primary Objective, and Acceptance criteria rewritten coherently: the hypothesis now names the pooled top-1 accuracy endpoint on the full image set (with rare-disease pathologies included in the pool), the Primary Objective matches that endpoint, and the rare-disease pooled subgroup is explicitly a pre-specified secondary endpoint rather than an implicit primary one.
• Item 2 — acceptance-criteria honest pass/fail. CIR §Per-specialty secondary analysis and §Endpoints not met together replace the prior "Both groups met … ✓ ACCEPTED" blanket (see Decision A).
• Item 3 — device version identity bridge. CIP and CIR both declare v1.1.0.0 with PRRC sign-off (Q4).
• Item 4 — formal sample-size justification. CIP §Sample size rewritten with MRMC power justification (see Decision K).
• Item 5 — "Risks: None" removed. CIP §Risks and Benefits rewritten as a proper risk enumeration with trace to R-TF-013-002 (see Decision M).
• Item 6 — MRMC Rank 11 framing. CIP §Scope §Nature and positioning and CIR §Research Title §Nature and positioning both state the Rank 11 / Pillar 3 §4.4 / MDR Article 2(48) posture explicitly (Q1).

Follow-ups to track

1. Trial-registration ID verification. NCT07428915 and EUPAS1000000910 should be verified against the external registries to confirm (i) the IDs exist, (ii) first-posted date, (iii) content of the public-facing results summary, and (iv) that the public-facing description is aligned with the Rank 11 / Pillar 3 §4.4 framing now in force. If either ID is invalid or was post-dated, a further CIR deviation entry should be added (D-03) documenting the registration status and timing with the date of actual registry posting. This sits outside the repository.
2. Publication Status citation. The prior "JMIR Dermatology Journal" claim is unverified. If the publication has been accepted, add the full citation (DOI, volume, issue, pages) to §Publication Status and confirm that the published framing matches the Rank 11 / Pillar 3 §4.4 positioning. If the publication was rejected or withdrawn, remove the §Publication Status section or rewrite it accordingly.
3. <StudyMetricsByTaskTable /> defaults. The CIR now renders the component bare (no taskLabels prop), relying on the component's defaults to display the BI_2024 performance claims. The component defaults should be audited to confirm they select studyId === "BI_2024" and present the expected task-level metric breakdown; if the default fallback shows all studies or a different study, a studyCode="BI_2024" prop should be added.
4. Per-pathology Wilson CIs. The per-pathology tables in CIR §Analysis report absolute differences but not per-pathology CIs; the CIP §Statistical analysis declares Wilson CIs as the method. Adding per-pathology Wilson CI columns is desirable but was deferred in this pass because the per-pathology analyses are already labelled exploratory; a future pass can emit the CIs from the locked analytics environment and append them to the table.
5. Ethics Committee Non-Applicability snippet — Spanish / Portuguese translation parity. Unchanged from SAN_2024 / MAN_2025 / PH_2024 follow-ups; English-only is currently acceptable for MDR CIP/CIR usage.
6. Data-driven acceptance-criteria renderer rollout. MAN_2025 introduced a pattern where the CIR-side AcceptanceCriteriaResultsTable.tsx reads thresholds from performanceClaims.ts. BI_2024 currently relies on <AcceptanceCriteriaTable displayMode="table" showAchievedValues={true} showStateOfArt={true} showUserGroup={true} showValueType={true} /> which is the same packages/ui data-driven component — no bespoke BI_2024 renderer was introduced. The acceptance-criteria pass/fail call in the CIR is carried by that component. No rollout action required for BI_2024.
7. CER consistency check. The CER (R-TF-015-003) should be checked to confirm that the BI_2024 entry in the Pillar 3 §4.4 supporting-evidence list is consistent with the reframed CIP/CIR and specifically that the Conclusions of the CER do not reintroduce the Maul 2024 or Liu 2016 citations as direct causal anchors for BI_2024. This was handled for MAN_2025 in the CER during the MAN_2025 pass; BI_2024 CER consistency should be re-confirmed.

Build verification

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

---

Second-pass closure — BSI re-audit and Celine framework review (2026-04-19)

After the full three-file rewrite above, the bsi-clinical-auditor and the Celine/Saray three-pillar framework (per .claude/agents/celine-clinical-consultant.md) were re-run against the rewritten files. Summary:

• BSI re-audit: CLOSED: ~12 prior findings | OPEN: 11 remaining. Three of the eleven were residual majors (primary-vs-secondary confirmatory ambiguity in the CIP acceptance structure; dermatologist "Pass" on n = 4 being regulatorily fragile; missing concrete R-TF-013-002 risk-row code for the automation-bias trace). Eight were minors (discordant-pair assumption missing from CIR sample-size summary, "approximately" on a Newcombe CI, "Endpoints not met" subsection title misleading, Publication Status still not aspirational enough, GDPR anonymisation-determination cite missing from CIP, Annex E Article 62/74/82 coverage thin, Annex E internal-audit pointer missing, Pillar 2 ICD-11-category count reintroduced in downstream §Introduction and §Discussion).
• Celine framework: FRAMEWORK ALIGNED: yes. Five minor polish items (one anchoring-wording variant on the patient-benefit chain; three Top-5 prioritised differential-view anchor-consistency items in the CIR §Material and methods and §Conclusions; one Pillar 2 / Pillar 1 triangulation sentence missing from Annex E §Applicability and scope).

What was applied in the closure pass

• CIP §Justification of acceptance thresholds rewritten to explicitly name the pooled endpoint as the primary confirmatory endpoint, and the per-specialty thresholds (PCP ≥ 10 pp / dermatologist ≥ 5 pp) and the rare-disease pooled endpoint as secondary confirmatory endpoints, with an explicit acknowledgement that the dermatologist stratum is under-powered and the per-specialty pass/fail call on the dermatologist stratum is supportive only. Mirrored in CIR §Material and methods → Acceptance criteria.
• CIR §Per-specialty secondary analysis table changed the dermatologist-row label from "Pass" to "Directional, supportive", with an accompanying paragraph stating the stratum is treated as a directional, supportive observation only — not as an independent confirmatory pass.
• CIR §Summary → Sample size augmented to name the discordant-pair-rate assumption ("an expected unaided baseline accuracy of approximately 50% and an expected aided accuracy of approximately 60%, corresponding to a discordant-pair rate consistent with state-of-the-art aided-HCP performance documented in R-TF-015-011") so the ≥ 80% power claim is reproducible.
• CIR §Primary Analysis — confirmatory Newcombe CI phrasing changed from "approximately 12.5–17.6 pp" to a reference to the locked analysis dataset with direction and magnitude consistent with the point estimate (removes the "approximately" on a pre-specified CI in regulatory prose).
• CIR §Analysis — "Endpoints not met" subsection renamed "Exploratory per-pathology comparisons that did not reach statistical significance" with a one-sentence preamble stating these were never elevated to endpoint status.
• CIR §Paediatric age-band subgroup (tinea corporis paragraph), §Conclusions and §Limitations (item 10) updated to cite the specific R-TF-013-002 risk rows R-DAG (device outputs a wrong result / HCP relies on the output) and R-HAX (HCP validates the wrong skin condition / misinterprets device output) as the automation-bias trace, and to state explicitly that the post-observation residual-risk re-evaluation does not trigger a new risk control at this time (existing controls — IFU warnings, Top-5 prioritised differential, clinician remains the decision-maker — apply; residual-risk determination is revisited if the paediatric tinea corporis signal is confirmed in a prospective or PMCF dataset). Mirrored in §Target-pathology exploratory description for the 1.77% overall and 8.08% GPP-specific PCP switch-to-incorrect proportions.
• CIR §Publication Status softened to aspirational: "The manufacturer intends to submit these results for peer-reviewed publication" (rather than "have been submitted"), with the full citation to be appended once the publication record is available.
• CIP §Informed Consent process → For image subjects augmented with a cite to the sponsor's data-protection file and to Article 29 Working Party Opinion 05/2014 on anonymisation techniques as the basis for the "non-personal data under GDPR" determination.
• Annex E Regulatory Authority Authorisation row expanded to explicitly address MDR Articles 62, 74 and 82 and to reconcile the NCT and EUPAS trial registrations with the "no regulatory-authority authorisation applicable" determination as a transparency measure.
• Annex E Audit Certificates row augmented with a pointer to the sponsor's internal-audit and quality-assurance procedures (GP-004 Internal Audits within the manufacturer's QMS) as the locus of the internal-monitoring-execution and data-verification records.
• CIR §Introduction and §Discussion → Regulatory positioning of this evidence simplified to remove the downstream repetitions of "346 ICD-11 categories at API level"; the full Pillar 2 scope statement is retained in §Nature and positioning only. Downstream repetitions now read "Pillar 2 (algorithm-level analytical performance) is evidenced independently at Clinical Evaluation level and is not the subject of this investigation".
• CIR §Discussion → Potential clinical-system implications phrase "the mechanism by which better primary-care diagnostic accuracy may translate into earlier correct treatment and more efficient referral pathways" rewritten as "sit on the surrogate leg of the surrogate-to-outcome chain characterised by the Pillar 1 literature; the outcome leg itself is neither measured nor inferred from this investigation" (removes the residual "translate into" causal verb that edged toward a direct accuracy→outcome claim).
• CIR §Material and methods → Objectives restated primary objective now includes "in its Top-5 prioritised differential-view configuration" to align verbatim with the Summary §Primary objective and the CIP.
• CIR §Discussion → Conclusions first sentence now names "the device's Top-5 prioritised differential view" (previously "the device") so the Pillar 3 UI anchor appears in the single most-quoted sentence of the document.
• Annex E §Applicability and scope augmented with a Pillar 2 / Pillar 1 triangulation sentence matching the CIP and CIR openings, so that Annex E read stand-alone still carries the full three-pillar framing.

Build verification (closure pass)

npx turbo run build --filter=qms passes after the closure-pass edits (2026-04-19).

State after closure pass

• All three rewritten files (r-tf-015-004.mdx, r-tf-015-006.mdx, r-tf-015-010.mdx) have been verified against both the BSI clinical-auditor framework (Erin Preiss / Nick) and the Celine/Saray three-pillar framework.
• The Celine framework verdict is FRAMEWORK ALIGNED: yes.
• The BSI clinical-auditor's three majors (primary/secondary confirmatory ambiguity; dermatologist Pass on n = 4; missing R-TF-013-002 risk-row codes) and all eight minors are closed.
• Follow-ups that remain (tracked elsewhere) are the ones that sit outside this repository: trial-registration verification with the external registries, full publication citation once submission is confirmed, and CER consistency cross-check for the reframed BI_2024 evidence (see §Follow-ups to track above).