How Relative Health Report builds its insights

1. What is this app, and what is it not?

Relative Health Report is a family health history tool. It helps you record and organise your family's health information, then share it with your doctor. It is not a diagnostic tool, not a substitute for genetic testing, and not a substitute for professional medical advice.

The insights surfaced in the app are guideline-anchored discussion points, not personal medical directives. Every clinical-sounding insight ends with a "talk to your doctor" prompt because that is the only safe path from a family-history pattern to a screening, treatment, or referral decision.

2. Where do our insights come from?

We draw on the same evidence base clinicians use. When you see an insight in the app, it traces back to one or more of these authoritative sources:

• U.S. Preventive Services Task Force (USPSTF) — Grade A/B recommendations for cancer risk assessment, statin prevention, and other primary-care screening decisions. Grade definitions.
• National Comprehensive Cancer Network (NCCN) — Genetic/Familial High-Risk Assessment guidelines for breast, ovarian, pancreatic, and colorectal cancer. Category 1 / 2A / 2B / 3 evidence labels.
• American College of Medical Genetics (ACMG) and the National Society of Genetic Counselors (NSGC) — practice guidelines for referral indications for cancer-predisposition assessment (Hampel et al. 2015) and the 2022 pedigree-nomenclature standard.
• U.S. Surgeon General / CDC Family Health History Initiative — the federal reference model for patient-collected family history.
• NIH National Human Genome Research Institute (NHGRI) — risk-stratification literature for common and hereditary conditions.

The research showing how strongly family history predicts disease risk is summarised on the science page.

3. How is AI used?

We use a large language model (currently Anthropic Claude) to read your family history and surface patterns that match the published guidelines above. The model is locked to a pinned version — we do not silently swap it for a newer one. When we do upgrade, we record it in the changelog (section 6).

Every clinical-sounding insight is paired with two trust artifacts:

• An evidence-grade chip (USPSTF A/B/C/D/I, NCCN Category 1–3, ACMG, or "Pattern-matched" for AI inferences that don't map cleanly to a single criterion). Tap the chip to see the underlying guideline.
• A provenance crumb back to the family-history data the insight was derived from (e.g., "Based on: 2 maternal aunts with breast cancer, age <50").

AI outputs can be incomplete or inaccurate. They are not medical advice, not a diagnosis, and must not be used to make medication, screening, or treatment decisions without a clinician. If an insight surprises you, take it to your doctor.

4. What do we deliberately not do?

• We do not diagnose disease. We surface patterns that match published referral and screening criteria and present them to you — and your doctor — for discussion.
• We do not generate personal risk numbers. Validated risk calculators (Tyrer-Cuzick, BOADICEA, MMRpro, PREMM5, Reynolds, ASCVD) belong in a clinician's hands. The report you hand to your doctor includes the inputs those tools expect.
• We do not order or interpret genetic tests. We help you document the results of tests you already have.
• We do not establish a clinician relationship. No one in the app is your doctor.
• We do not share your data with insurers, advertisers, or any third party you have not explicitly authorised.

5. How do we grade insights?

Every clinical-sounding insight carries an evidence-grade chip. The grading vocabulary:

• USPSTF A / B — Task Force recommends the service; substantial or moderate net benefit at high certainty.
• USPSTF C — offer for selected patients, small net benefit.
• USPSTF D — Task Force recommends against the service.
• USPSTF I — insufficient evidence to assess the balance.
• NCCN Category 1 / 2A / 2B / 3 — high consensus high evidence, through major disagreement.
• ACMG — referral-indication criteria from the American College of Medical Genetics / National Society of Genetic Counselors practice guideline.
• Pattern-matched — the insight derives from a pattern in your family history that doesn't precisely match a published criterion. Lower confidence; review with your clinician.

6. How do we handle versioning and model changes?

When we change the AI model, the guideline library, or the insight-generation logic, we record it here. We do not silently update clinical content.

Current versions:

• AI model: Anthropic Claude (pinned version documented in each generated insight payload).
• Guideline library: USPSTF (2024 update), NCCN Genetic/Familial High-Risk Assessment v2025.1, ACMG Hampel 2015, NSGC pedigree nomenclature 2022.
• ICD-10 Z-code mapping: 2025 fiscal year codes; mapping table reviewed on each major release.

7. How do GINA and your data privacy work?

You own your data. We don't sell it, we don't license it to insurers or advertisers, and we don't use it for any purpose you haven't explicitly authorised.

One legal nuance you should know about before sharing family history with any clinician:

This is a consequence of where the data goes once it leaves the app — not of how the app handles your data. See our security white paper for exactly what we can and cannot see.

8. Do we have conflicts of interest?

We do not receive payment, equity, or kickbacks from any lab, genetic-testing company, pharmaceutical company, or insurer to surface particular insights or recommend particular products. The clinical content in the app reflects the public guidelines listed in section 2 — nothing else.

If a clinical advisor reviews app content, we disclose their name and relationship here.