Transparency · CHAI Applied Model Card
Insulin Resistance Detection
CHAI Applied Model Card (v0.1) for the Insulin Resistance Detection pattern in Zenlo Labs. This document is not CHAI-certified or CHAI-endorsed.
Header
Summary
The Insulin Resistance Detection pattern in Zenlo Labs combines a deterministic HOMA-IR computation — derived from fasting glucose and fasting insulin — with an LLM-assisted clinical narrative to flag insulin resistance patterns in laboratory panels. It is intended for licensed functional-medicine physicians as clinical decision support: the tool surfaces a pattern flag and explanatory context so a physician can review results alongside the full clinical picture. It is supportive, not diagnostic. HOMA-IR provides a reproducible numeric anchor; the narrative layer synthesizes related biomarkers and contextual labs into physician-facing language. Outputs are meant to prompt clinical review, not to replace physician judgment or definitive metabolic testing.
Uses & Directions
Intended use
Clinical decision support (CDS) for licensed physicians evaluating metabolic patterns in adult laboratory panels.
Primary users
Licensed functional-medicine physicians and similarly qualified clinicians.
How to use
Review pattern output alongside the full clinical picture, patient history, and corroborating tests. Clinical use requires physician authorization within the Zenlo Labs workflow.
Target population
Adults aged 18 and older in the United States.
Out of scope
- Direct patient use without physician oversight
- Pediatric populations
- Standalone diagnosis or treatment decisions
- Type 1 diabetes management
Warnings
Clinical risk level
Low — supportive tool; the treating physician retains full clinical judgment and responsibility.
Known limitations
- HOMA-IR is validated for use in the fasting state only.
- Population-specific HOMA-IR thresholds vary in the literature.
- Not a substitute for oral glucose tolerance testing (OGTT) or hyperinsulinemic euglycemic clamp studies.
- Detected prevalence varies across demographic subgroups (see Key Metrics fairness table).
Bias note
Descriptive prevalence of detected patterns differs across demographic groups, as shown transparently in Key Metrics. These are population descriptive rates from an independent audit, not model error rates.
Trust Ingredients
AI system facts
- Deterministic HOMA-IR detector plus Claude Haiku 4.5 for narrative synthesis
- Primary inputs: fasting glucose and fasting insulin, with contextual laboratory values when available
- Output: pattern flag plus physician-facing explanation
Security & compliance
- Anthropic Business Associate Agreement with zero-data-retention configuration
- HIPAA-aligned design; no patient data used for model training
Ongoing maintenance
Versioned, transparent, and reproducible via a public independent audit harness (see Resources).
Transparency
Self-funded development; no third-party sponsor for this pattern card.
Key Metrics
Usefulness / Efficacy
Detector accuracy from a separate 5-model benchmark study — not from the NHANES descriptive audit below. Source: 5-model benchmark, medRxiv MEDRXIV/2026/346284.
Fairness / Equity
Overall prevalence detected: 23.1% (1,387 / 5,992).
By sex
| Subgroup | n | Prevalence detected |
|---|---|---|
| Female | 3,105 | 22.3% |
| Male | 2,887 | 24.1% |
By age group
| Subgroup | n | Prevalence detected |
|---|---|---|
| 18–39 | 2,226 | 18.7% |
| 40–59 | 1,865 | 25.3% |
| 60+ | 1,901 | 26.3% |
By race / ethnicity
| Subgroup | n | Prevalence detected |
|---|---|---|
| Mexican-American | 1,064 | 25.7% |
| Other Hispanic | 798 | 26.6% |
| Non-Hispanic White | 1,914 | 23.4% |
| Non-Hispanic Black | 1,265 | 20.4% |
| Non-Hispanic Asian | 726 | 19.8% |
| Other / Multiracial | 225 | 23.1% |
Prevalence rises with age and varies by demographic group, consistent with known epidemiology. These are descriptive population rates from the audit harness — not model error rates or benchmark accuracy metrics. Source: Independent NHANES 2015–2016 audit harness (N=5,992, commit c10afe8).
Safety / Reliability
- Supportive-only; not intended as a standalone diagnostic
- Physician authorization required before clinical use
- Deterministic HOMA-IR component is reproducible: identical inputs produce identical numeric output
Resources
- medRxiv preprint MEDRXIV/2026/346284 — 5-model benchmark (F1 and related detector metrics)
- JAMIA Open submission JAMIO-2026-0120 (under review)
- Independent audit harness: github.com/dimashibakov/zenlo-audit — reproducibility manifest, NHANES 2015–2016 cycle, harness commit c10afe8
Note: The mention or sharing of any examples, products, organizations, or individuals does not indicate any endorsement by the Coalition for Health AI (CHAI). Zenlo Labs is a clinical decision support tool intended exclusively for use by licensed healthcare providers — not a substitute for professional medical judgment, and not intended for direct patient use.