IYN Clinical Lab

1. Manifesto: Why Deterministic Rule Engines Must Anchor Clinical Oncology

“In general web copy or creative writing, a 95% accurate language model is impressive. In gynecologic oncology, a model that guesses 5% of its thresholds is a clinical hazard.”

IYN Clinical Lab is an independent, bootstrapped clinical informatics research project. We do not believe the future of oncology decision support lies in prompting massive, general-purpose foundation models with unstructured medical charts and hoping their probabilistic token predictions align with regulatory labels.

Modern gynecologic oncology—spanning epithelial ovarian, fallopian tube, primary peritoneal, endometrial, and cervical malignancies—is governed by strict quantitative laboratory cutoffs, sequential diagnostic hierarchies, and subgroup-restricted survival endpoints. When selecting a first-line maintenance regimen for advanced ovarian cancer, the difference between a progression-free survival hazard ratio of 0.33 (profound clinical benefit) and 1.00 (zero benefit with substantial hematologic toxicity) depends entirely on a verified Homologous Recombination Deficiency (HRD) genomic instability threshold.

Our engineering principle is simple: probabilistic language models should never be trusted to recall or interpolate hard clinical facts. Clinical eligibility, biomarker cutoffs, staging shifts, and pharmacokinetic dose adjustments must be executed by transparent, deterministic logic anchored directly to peer-reviewed primary manuscripts.

2. Four Structural Failure Modes of Unconstrained LLMs in Gynecologic Oncology

When ungrounded generative models summarize oncology literature or evaluate pathology reports, four reproducible failure modes consistently emerge:

1. Assay Platform & Threshold Conflation: Molecular biomarkers are not interchangeable boolean flags. In high-grade serous ovarian carcinoma, “HRD positivity” is defined by distinct platform-specific cutoffs: a Genomic Instability Score (GIS) ≥ 42 on the Myriad myChoice CDx assay (used in PAOLA-1 and PRIMA), versus a Loss of Heterozygosity (LOH) score ≥ 16% on FoundationOne CDx. Similarly, in platinum-resistant ovarian cancer, eligibility for Mirvetuximab Soravtansine requires Folate Receptor-alpha (FRα) expression scored specifically via the VENTANA FOLR1 (FOLR1-2.1) RxDx Assay at ≥ 75% of viable tumor cells with ≥ 2+ membrane staining intensity (PS2+)—not merely “any positive IHC staining.”
2. Subgroup Evidence Evaporation: In dense clinical summaries, critical negative subgroup outcomes are routinely omitted. In the Phase III PAOLA-1 trial, adding maintenance Olaparib to Bevacizumab improved median progression-free survival to 37.2 vs. 17.7 months in HRD-positive tumors (HR: 0.33), but yielded an exact Hazard Ratio of 1.00 (95% CI: 0.75–1.35) in HR-proficient (HRp) disease. By contrast, single-agent Niraparib in the PRIMA trial demonstrated statistically significant activity even in the HR-proficient cohort (HR: 0.68; 95% CI: 0.49–0.94). Conflating these two trials leads directly to inappropriate prescribing.
3. Hierarchical Molecular Classifier Inversion: Approximately 3% to 6% of endometrial carcinomas harbor more than one molecular alteration (e.g., concurrent POLE exonuclease domain mutation and abnormal p53 overexpression, or dMMR plus abnormal p53). Probabilistic models heavily weight the keyword “p53 abnormal” and misclassify these tumors as high-risk, recommending aggressive adjuvant chemoradiotherapy. In reality, validated clinical genomics (ProMisE / TransPORTEC) proves that POLE-ultramutated and dMMR phenotypes are upstream drivers, and passenger TP53 variants in these tumors do not worsen prognosis.
4. Anatomical & Pharmacokinetic Contraindication Blindness: Text-similarity models lack causal verification across pelvic anatomy, organ clearance, and metabolic pathways. Recommending Bevacizumab in locally advanced cervical carcinoma with transmural rectovaginal septum invasion invites catastrophic rectovaginal fistula formation. Prescribing full-dose Niraparib (300 mg QD) to a 62 kg patient without applying the mandatory individualized starting dose rule (< 77 kg or platelets < 150,000/μL → 200 mg QD) causes severe, preventable Grade 3/4 thrombocytopenia.

3. Deterministic Diagnostic Hierarchy: Endometrial Cancer Molecular Subtyping (ProMisE & FIGO 2023)

To prevent classifier inversion in endometrial carcinoma, our eligibility logic enforces the strict sequential evaluation order established by the Proactive Molecular Risk Classifier for Endometrial Cancer (ProMisE), ESGO/ESTRO/ESP guidelines, and the revised 2023 FIGO staging system. Evaluation must proceed strictly from Step 1 to Step 3:

1. Step 1 — Evaluate POLE Exonuclease Domain Mutation Status (NGS / Sanger Sequencing):
   • Condition: Presence of a verified pathogenic variant in the exonuclease domain of POLE (codons 268–471; most commonly P286R, V411L, S297F, A456P, or S459F).
   • Deterministic Output: Classify as POLEmut (Ultramutated). Stop further molecular risk escalation.
   • Multiple-Classifier Rule: Even if mismatch repair immunohistochemistry is deficient or p53 IHC shows aberrant overexpression/null pattern, a verified pathogenic POLE exonuclease domain mutation overrides downstream markers (classified strictly as POLEmut; reference: León-Castillo et al., J Pathol 2020; PMID: 31829442).
   • FIGO 2023 Stage Shift: In tumors confined to the uterine corpus or with cervical extension, pathogenic POLEmut status downstages the tumor to Stage IAm_POLEmut regardless of lymphovascular space invasion (LVSI) or histological grade, supporting treatment de-escalation.
2. Step 2 — If POLE Wild-Type, Evaluate Mismatch Repair (MMR) Status via 4-Antibody IHC or MSI PCR:
   • Condition: Complete loss of nuclear expression of at least one MMR protein (MLH1/PMS2 heterodimer or MSH2/MSH6 heterodimer) OR high microsatellite instability (MSI-H).
   • Deterministic Output: Classify as dMMR (Mismatch Repair Deficient / Hypermutated). Stop further escalation.
   • Multiple-Classifier Rule: If a tumor is POLE wild-type and exhibits both dMMR and aberrant p53 staining (dMMR + p53abn), the p53 abnormality is treated as a secondary passenger event induced by hypermutation; the tumor is classified as dMMR.
   • Therapeutic Implication: High sensitivity to front-line anti-PD-1 checkpoint blockade combined with carboplatin/paclitaxel (RUBY Part 1 dMMR PFS HR: 0.28; NRG-GY018 dMMR PFS HR: 0.30).
3. Step 3 — If POLE Wild-Type and MMR-Proficient (pMMR), Evaluate p53 Status (IHC / TP53 Sequencing):
   • Condition A (Abnormal / Mutant-Pattern p53): Overexpression (strong nuclear staining in > 80% of tumor cells), complete absence (null pattern with positive internal control), or unequivocal cytoplasmic staining.
     • Deterministic Output: Classify as p53abn (Copy-Number High / Serous-like).
     • FIGO 2023 Stage Shift: Any myoinvasive endometrial carcinoma with p53abn confined to the uterine corpus (with or without cervical invasion) is upstaged directly to Stage IICm_p53abn, mandating adjuvant systemic therapy consideration.
   • Condition B (Wild-Type Pattern p53): Heterogeneous nuclear staining of variable intensity (1% to 80%).
     • Deterministic Output: Classify as NSMP (No Specific Molecular Profile / Copy-Number Low). Risk stratification within NSMP is subsequently refined by Estrogen Receptor (ER) expression status, histological grade, and substantial LVSI (≥ 5 vessels).

4. Biomarker Cutoff & Companion Diagnostic Reference Matrix

The structured matrix below defines the exact diagnostic assays, quantitative positivity thresholds, and corresponding regulatory-approved targeted or immunotherapeutic regimens across gynecologic malignancies:

5. Audited Phase III Clinical Trials Corpus (10 Seminal Publications)

Every clinical trial specification, sample size, hazard ratio, confidence interval, and p-value below has been audited directly against the primary peer-reviewed publication in The New England Journal of Medicine (NEJM) or The Lancet, alongside subsequent peer-reviewed final survival analyses in the Journal of Clinical Oncology (JCO) and The Lancet Oncology.

Part A: Epithelial Ovarian, Fallopian Tube & Primary Peritoneal Cancer (Trials 1–4)

Part B: Endometrial Carcinoma (Trials 5–7)

Part C: Cervical Carcinoma (Trials 8–10)

6. Deterministic Pharmacovigilance, Dosing Formulas & Contraindication Rules

Safe deployment of targeted agents and antibody-drug conjugates (ADCs) in gynecologic oncology requires strict pre-treatment checks across anatomy, anthropometrics, organ clearance, and enzyme interactions.

7. Open Deterministic Evidence & Eligibility Schema (v0.1 Reference Specification)

Rather than relying on opaque prompt engineering, IYN encodes each clinical trial and guideline rule as a deterministic, machine-verifiable JSON structure. Below is a production excerpt demonstrating how assay thresholds, PICO provenance, and negative subgroup constraints are represented for zero-hallucination evaluation:

{
  "$schema": "https://iyn.one/schemas/clinical-evidence-node-v0.1.json",
  "rule_id": "GYNONC-OV-1L-MAINT-PAOLA1",
  "disease_ontology": {
    "snomed_ct": "254837009",
    "label": "High-grade serous or endometrioid epithelial ovarian, fallopian tube, or primary peritoneal carcinoma",
    "figo_stages": ["III", "IV"]
  },
  "provenance": {
    "trial_acronym": "PAOLA-1 / ENGOT-ov25",
    "nct_id": "NCT02477644",
    "primary_pmid": "31851799",
    "followup_os_pmid": "37211045",
    "doi": "10.1056/NEJMoa1911361"
  },
  "deterministic_eligibility_gate": {
    "prior_therapy_requirement": "CR_OR_PR_TO_1L_PLATINUM_TAXANE_PLUS_BEVACIZUMAB",
    "biomarker_logic": {
      "operator": "OR",
      "conditions": [
        {
          "biomarker": "BRCA1_OR_BRCA2",
          "variant_classifications": ["PATHOGENIC", "LIKELY_PATHOGENIC"]
        },
        {
          "biomarker": "HRD_GENOMIC_INSTABILITY_SCORE",
          "assay": "Myriad myChoice CDx",
          "comparator": ">=",
          "threshold": 42
        }
      ]
    }
  },
  "subgroup_outcomes": {
    "HRD_POSITIVE": {
      "recommendation_status": "INDICATED_CATEGORY_1",
      "regimen": "Olaparib 300 mg PO BID (up to 24 mos) + Bevacizumab 15 mg/kg IV q3w (15 mos total)",
      "median_pfs_months": { "intervention": 37.2, "control": 17.7 },
      "pfs_hazard_ratio": 0.33,
      "pfs_95_ci": [0.25, 0.45],
      "five_year_os_hazard_ratio": 0.62,
      "five_year_os_95_ci": [0.45, 0.85]
    },
    "HR_PROFICIENT_GIS_LT_42": {
      "recommendation_status": "CONTRAINDICATED_NO_EFFICACY",
      "median_pfs_months": { "intervention": 16.6, "control": 16.2 },
      "pfs_hazard_ratio": 1.00,
      "pfs_95_ci": [0.75, 1.35],
      "clinical_warning": "Zero progression-free survival benefit in HR-proficient disease; do not add Olaparib to Bevacizumab maintenance."
    }
  }
}

8. Lab Status & Research Roadmap

Stage: Independent Pre-Seed R&D (Self-Funded).

Infrastructure: Hosted on hardened Linux nodes running Caddy web server behind Cloudflare Zero Trust & WireGuard mesh. Zero third-party JavaScript, zero external fonts, zero analytics trackers, and zero cookies.

Current Milestones:

• [x] Milestone 1: Audit and structure 10 practice-defining Phase III gynecologic oncology trials (NEJM, Lancet, JCO) with complete PICO & subgroup confidence bounds.
• [x] Milestone 2: Codify the deterministic ProMisE / FIGO 2023 endometrial cancer molecular classification decision tree (resolving multiple-classifier conflicts).
• [x] Milestone 3: Encode quantitative dosing & safety gates (Niraparib ISD, Mirvetuximab AIBW, Olaparib CrCl/CYP3A4 reductions, Bevacizumab fistula anatomical exclusion).
• [ ] Milestone 4 (In Progress): Publish open, downloadable JSON-Schema and HL7 FHIR R4 EvidenceVariable / PlanDefinition bundles for all 10 trials.
• [ ] Milestone 5 (Planned): Release an open-source, client-side verifiable CLI and WebAssembly rule evaluator for academic gynecologic oncology tumor boards.

9. Contact & Academic Collaboration

We welcome correspondence from gynecologic oncologists, molecular pathologists, biostatisticians, and clinical informatics researchers interested in auditing deterministic oncology rule sets or collaborating on structured trial extraction.

Direct Electronic Mail: contact@iyn.one

No commercial forms, no tracking pixels, and no marketing lists—direct technical and academic correspondence only.