Reference ranges & thresholds
BMI 23/25 · IDF waist 90/80 · sex-specific low-HDL cuts · earlier diabetes and CAC screening ages · the atherogenic-dyslipidemia pattern · Lp(a) test-once. LOINC-coded, FHIR-mappable.
Open Masala.
The open, ancestry-adjusted
health reference layer.
South Asian screening thresholds — lower BMI cutoffs, earlier diabetes screening, IDF waist limits, Lp(a) prompts — already exist. They're just scattered across guideline PDFs and hundreds of papers, never assembled anywhere you can query. Open Masala pulls them into one machine-readable dataset — every value LOINC-coded, evidence-graded, and cited to its source.
Free & open
CC-BY 4.0
LOINC + FHIR-shaped
Every row cited
Built for LLMs
Why this exists
The numbers exist. The dataset doesn't.
The WHO Asian BMI cutoffs live in a 2004 consultation, the IDF waist limits in a consensus PDF, the HbA1c under-read in a cohort paper. The South Asian thresholds are real and often guideline-endorsed — they've just never been pulled into one structured, cited, downloadable table. That's the gap Open Masala fills. (We synthesize the published thresholds; we don't re-host the gated cohort data behind them.)
What's in it
One row, and everything you need to trust it.
The core artifact is a table of ancestry-stratified reference ranges and screening thresholds. Every row pairs the South Asian value with the general-population value it replaces, so the delta is visible at a glance — and carries a provenance tier telling you exactly what kind of claim it is.
Diabetes screening trigger — BMI
LOINC-coded · kg/m²
South Asian
≥ 23
General population
≥ 25
Sources: ADA Standards of Care 2024 (DOI 10.2337/dc24-S002) · Araneta & Kanaya,
Diabetes Care 2015 (PMID 25665815).
Overclaim guard: a screening trigger is not a diagnosis.
A cited evidence corpus
100+ curated citations — named guidelines and named studies, each with an evidence grade and a live, validated link. The reference values point back into it.
Machine-readable honesty
Every row carries an overclaim guard — the claim you must not make from it. Exactly what an LLM feature needs to keep from stating an ancestry claim with false confidence.
Credibility
Curated with a viewpoint — and honest about its edges.
This is not a scrape. It's an editorial dataset with a rating system, and we'd rather tell you where it's thin than pretend it's complete. Every row is tagged with one of three provenance tiers.
guideline-endorsed
A named body recommends this exact cutoff for this population — WHO Asian BMI, IDF waist, ADA South Asian screening. Highest confidence.
study-derived
A real, cited effect estimate from the literature that hasn't been codified into a guideline cutoff — like the HbA1c under-read. Useful, but never presented as settled practice.
contested-deprecated
An adjustment the field is moving away from — like the eGFR race coefficient, retired in 2021. Included precisely so we can tell you not to apply it.
Who it's for
Four audiences, one dataset.
Health teams
A product team that wants ancestry-aware calibration in their own experience —
a device maker, a lab, a care platform. FHIR-shaped and LOINC-coded, so it drops
into a health stack in days, not quarters.
Developers
One
git clone or package install and you have every threshold as
structured data, with helper code and a data dictionary. No parsing PDFs.
LLMs & agents
Retrievable and tool-callable — an
llms.txt surface and an MCP server
you can add to Claude in seconds — so an agent quotes a cited South Asian
reference range instead of hallucinating one.
Researchers
A citable, DOI-minted, versioned release. Every row traces to its source, so it's
safe to cite and reproduce — a synthesis contribution, not a primary-data claim.
Access
Open by default. Pick your front door.
The data is free and live across the channels each audience actually looks in — downloadable, ML-ready, citable, and agent-callable.
Git repo ↗
Canonical home — tables, schema, and data dictionary. JSON, CSV, and a FHIR bundle, CC-BY-4.0.
Live
Personal advisor ↗
Run it as your own ancestry-aware health advisor — your data stays on your machine.
Live
Get started
Three ways in — each a copy-paste.
Get the data
Clone the repo, or pull it straight from Hugging Face.
git clone https://github.com/masalahealth/open-masala
# …or in Python
from datasets import load_dataset
ds = load_dataset("masalahealthco/open-masala")
Use it in an agent
Add the MCP server to Claude Desktop (or any MCP client), then ask about a value and
get a cited, ancestry-adjusted answer. Drop this into
claude_desktop_config.json:
{
"mcpServers": {
"open-masala": { "command": "uvx", "args": ["open-masala-mcp"] }
}
}
Cite it
Archived on Zenodo with a DOI — safe to cite and reproduce. Please also cite the
primary source named on each row.
DOI: 10.5281/zenodo.21300418 · CC-BY-4.0
Full docs, the FHIR bundle, and a run-it-yourself health advisor live in the repository.
Beyond the data
Turn it into your own health advisor.
Point Claude at your own MyChart export with the Open Masala MCP and you get a personal, ancestry-aware advisor — it reads your labs, tracks trends across years, and preps you for a sharper doctor visit, your data staying on your machine. Three steps: pull your records, build your profile, ask.
Read the guidePartner with us
Building something that should know about South Asian bodies?
If you're a product, research, or clinical team and ancestry-aware calibration would make your work better, tell us what you'd want from the dataset. We'll prioritize the channels and coverage that unblock real integrations.
We'll only email you about the dataset. Open Masala is a reference dataset, not a medical device, and not a substitute for clinical judgment.