How to map drugs to indications — and why open data gets it wrong

Why can't you map drugs to diseases with open data?

Our first attempt was the obvious one. The National Library of Medicine publishes drug–disease relationships through MED-RT, exposed in RxClass and RxNav. It's free, it's keyed to RxNorm, and it's maintained by people who know what they're doing. We wired it up, pointed it at real prescriptions, and the answers were wrong in ways no amount of cleanup could fix.

That's not a knock on NLM. MED-RT answers a different question — what is this drug known to treat? — and answers it as a flat list. Analytics needs what does the patient behind this fill have, and how often? Three gaps separate the two:

• Coded to the target, not the patient. A muscle relaxant maps to muscle spasm. The claim that goes with the fill says low back pain.
• No weights. Every relationship counts the same. A use that drives a third of volume and one that drives a rounding error look identical.
• Label-bound. Off-label use — often most of a drug's volume — is largely absent.

Cyclobenzaprine: the drug treats spasm. The claim says back pain.

MED-RT lists seven conditions cyclobenzaprine may treat:

Every one describes what the drug acts on — or, in tetanus, a use almost no outpatient prescription is for. None is the diagnosis a payer sees next to the fill. Weighted equally, tetanus gets the same 14% as everything else. Here is how we map the 10 mg tablet, with the billable code each use lands on:

MED-RT relationships retrieved from the NLM RxClass API, September 2026.

The same pattern runs through every symptom-relief drug. For ondansetron, MED-RT lists two relationships — postoperative nausea and alcoholism. In practice a quarter of ondansetron is written for nausea with no recorded cause, a fifth for gastroenteritis, and meaningful shares for pregnancy, chemotherapy and migraine. In each case the claim carries the underlying condition, not the nausea. Where no cause is recorded, we say so rather than invent one: a suspecting model that assigns cancer to every antiemetic fill is worse than one that admits it doesn't know.

Gabapentin: why equally weighted indications are useless

Cyclobenzaprine shows the coding problem. Gabapentin shows the weighting problem. MED-RT lists four uses — postherpetic neuralgia, partial seizures, bipolar disorder and phobic disorders. Consume them as-is and each gets a quarter of the volume. In practice those four account for about 11% of gabapentin prescribing; the drug's real footprint is neuropathic and radicular pain, almost all of it off-label.

Why build an indication ontology before mapping to ICD-10?

The tempting shortcut is to map drugs straight to ICD-10. ICD-10 is a billing classification: it forces every idea into whatever code a claim would carry, at whatever specificity the code set happens to offer. “Stroke prevention in atrial fibrillation” becomes an atrial fibrillation code and the prevention intent disappears. “Muscle spasm in acute low back pain” becomes either a spasm code or a back pain code — pick one, lose the other.

So the foundation is our own vocabulary — the TwinFyRx Indication — written the way a prescriber or pharmacist would say it, and coded at the level of detail the source actually supports. Each concept separates the parts a code collapses:

Patient has is the field that matters most: it is the diagnosis that would sit on the claim for that fill, which is what suspecting and claims matching need. Concepts can also carry a qualifier (first-line, adjunctive, palliative), up to three associated conditions, and a parent, so “painful diabetic neuropathy” and “postherpetic neuralgia” both roll up to “neuropathic pain, any cause.”

Keeping the ontology ours is what keeps us flexible. When ICD-10 revises a code, when a buyer wants SNOMED instead, or when we find two concepts that should be one, we change one mapping — not thousands of drug rows. The drug estimates never have to be redone because a code set moved.

How does it map to ICD-10, SNOMED CT and MeSH?

Nobody builds on our ontology — nor should they have to. Your claims speak ICD-10; your clinical systems speak SNOMED CT; your literature pipelines speak MeSH. So the last step translates every concept into a clinical condition layer built on AHRQ's Clinical Classifications Software Refined (CCSR), and from there into standard codes. Every ICD-10-CM code in our drug-level estimates is billable; where a source only supports a category, we roll it down to the unspecified child rather than publish a code a claim can't carry.

The published grain is the drug: every RxNorm clinical and branded drug carries its indication shares (drug_indication), and drug_icd10_estimate rolls them to billable codes. You join on drug_id and never have to learn our vocabulary unless you want to.

How is the mapping built?

We won't publish the recipe, but the shape is simple: a multi-stage, AI-assisted research workflow with domain experts deciding what ships. Five stages, each auditable, each result kept as an immutable, versioned run rather than overwritten — and every release tested against real-world prescription data before it is published.

1. 01

   Research

   Frontier AI models, run as a structured research workflow, work through each drug the way a clinical pharmacist would — drawing on FDA labeling, guidelines and the published real-world utilization literature — and record how it is actually used, on-label and off.

2. 02

   Normalize

   Express every use as a TwinFyRx Indication. Reuse an existing concept before minting a new one, so ten drugs for the same job share one concept.

3. 03

   Curate

   Review the vocabulary chapter by chapter: merge duplicates, build parent–child structure. Every change is logged and reversible.

4. 04

   Estimate

   Weight each drug's uses by share of prescriptions — per drug form, or per product line where one form serves different markets.

5. 05

   Translate

   Code each concept's conditions once — to a clinical condition group and billable ICD-10-CM — then publish at the drug (RxNorm SCD/SBD) grain.

Two design choices matter more than the rest. First, estimation happens at the level where use actually differs — usually the drug form, but per product line where one form serves different markets. Finasteride 1 mg is a hair-loss drug (91% male pattern hair loss); the 5 mg tablet is a prostate drug (77% BPH). Mounjaro and Zepbound are the same tirzepatide pen — one is 87% type 2 diabetes, the other 70% obesity. Second, the vocabulary is shared across the whole catalog, so a new drug inherits concepts its therapeutic neighbors already use instead of inventing near-duplicates.

How is it validated?

Indication shares are an estimate, and we treat them like one. Nothing publishes until it clears four gates:

Overlap is measured at the condition-group level on drug forms with at least 100 MEPS fills. The 0.81 ceiling compares two MEPS survey periods with each other — the agreement any method can expect given survey noise.

MEPS has its limits — it's self-reported, it under-sees specialty and physician-administered drugs, and its condition coding is coarse — so it's a floor for agreement, not a ceiling for truth. That is why it is one gate of four, and why data-quality results for the published tables live on our public data quality page.

Explore the indication hierarchy

Browse the published vocabulary by body system and condition, open a concept to see its structure and ICD-10 landing codes, and see the leading drugs for it — or search for a drug to find the indications it leads.

Use it: tables and API

The indication layer ships as five tables — drug_indication, indication, indication_condition, condition, condition_icd10 — plus the derived drug_icd10_estimate. For one drug at a time, call the API:

{
  "drug": { "drug_id": 828348, "drug_name": "cyclobenzaprine hydrochloride 10 MG Oral Tablet" },
  "indication_count": 12,
  "indications": [
    {
      "indication_name": "Muscle spasm in acute low back pain",
      "pct_volume": 35.0,
      "use_type": "on_label",
      "relationship": "treats_symptom_of",
      "primary_condition": "acute low back pain",
      "target": "muscle spasm",
      "conditions": [{ "condition_name": "Low back pain",
                       "icd10_codes": [{ "code": "M54.50", "is_billable": true }] }]
    },
    …
  ]
}

Full field reference on the data model page and in the API reference.

Frequently asked questions

Sources

• MED-RT drug–disease relationships via the NLM RxClass API, retrieved September 2026.
• Clinical Classifications Software Refined (CCSR) for ICD-10-CM Diagnoses. Healthcare Cost and Utilization Project (HCUP). Agency for Healthcare Research and Quality, Rockville, MD. hcup-us.ahrq.gov. TwinFyRx's condition layer is adapted from CCSR; it is not produced or endorsed by AHRQ.
• Medical Expenditure Panel Survey (MEPS), Prescribed Medicines and Medical Conditions files. Agency for Healthcare Research and Quality. meps.ahrq.gov. Used for statistical benchmarking only.
• ICD-10-CM, Centers for Medicare & Medicaid Services / National Center for Health Statistics.