How We Built Our Knowledge Base: The Research Behind TranslateMD

When we set out to build TranslateMD, we made a decision early on that has shaped everything since: medical code translation should not be guesswork.

AI language models are impressive. They can summarize, explain, and contextualize medical documents better than most tools available today. But ask an AI to map a German ICD-10-GM code to its US ICD-10-CM equivalent, and you’ll get a confident answer that may be subtly wrong in a way that costs a patient a claim denial — or worse, misleads a clinician.

We decided the core of TranslateMD would be a curated, validated knowledge base where every mapping is either confirmed by authoritative sources or explicitly flagged as an approximation. Here’s how we built it.

Why Curated Mapping for Medical Codes

Medical coding is not a fuzzy problem. When a German discharge summary contains I10.00 (essential hypertension, no hypertensive crisis), the correct US equivalent is I10. Not a range of possibilities — that specific code, with an annotation that the crisis status was lost in translation.

When a dental record from Japan uses FDI notation #26 for a tooth, the correct US Universal Notation equivalent is #14. There is no ambiguity. The tooth is the tooth.

AI inference introduces probabilistic error into a domain that has exact answers most of the time. Our evaluation data confirms that a well-built knowledge base consistently outperforms AI-only approaches on both standard code pairs and challenging edge cases — rare codes, partial matches, and system-specific extensions that don’t cleanly map to any equivalent.

For the small percentage of cases where no curated mapping exists, we do use AI-augmented approaches. But we always know which category a given mapping falls into, and we surface that information to the clinician. The goal is transparency, not false confidence.

How the Knowledge Base Is Organized

The TranslateMD knowledge base is organized in layers, each serving a distinct purpose in the translation workflow.

Country-Level Context

Every healthcare system starts with an understanding of how that country’s medical record ecosystem works. Not just the code sets, but the clinical context: how diagnoses are documented, what coding conventions are standard, which system-specific features exist that have no equivalent elsewhere.

Germany’s ICD-10-GM adds an exclamation mark suffix (!) to optional secondary codes — a syntactic convention that doesn’t exist in the WHO base or any other national variant. Japan’s medical system integrates Kampo (traditional herbal medicine) through a separate coding track that sits alongside conventional diagnosis coding. India’s records frequently mix WHO ICD-10 with AYUSH (Ayurveda, Yoga, Unani, Siddha, Homeopathy) documentation in ways that require context to interpret correctly.

Without understanding these system-level conventions, code mapping is unreliable.

Code-to-Code Crosswalks

The next layer covers the direct mappings between code systems: ICD-10-CM to ICD-10-GM, FDI dental notation to Universal/ADA notation, ATC drug codes to NDC/PZN identifiers, and so on.

These crosswalks are built from primary sources: official mapping tables from national health authorities, published crosswalk documentation, and clinical coding standards. Where official mappings exist, we use them. Where they don’t — and in many cases they don’t — we document the closest valid equivalent and the nature of the gap.

Every mapping has a confidence type: exact (the codes are definitionally equivalent), approximate (the target code is the best available equivalent with some specificity loss), or unmapped (no valid equivalent exists). This classification is surfaced in every translation output.

Translation Corridor Depth

Different country pairs have different translation challenges. The Germany-to-US corridor, for example, has unusually high complexity around drug safety alerts. Several medications that are standard German formulary drugs — metamizole (Novalgin), nimesulide — are either banned or strictly controlled in the US. A German prescription that a US pharmacist might encounter needs an immediate safety context, not just a code conversion.

The India-to-US corridor has different complexity: a high volume of generic drug names that don’t correspond to US brand names, and a large AYUSH documentation component that US providers may have never encountered. The knowledge base addresses these practical translation challenges with clinical depth specific to each country pair.

Notation Systems and Medical Terminology

Some coding systems are shared across countries but implemented differently. Dental notation is the clearest example: the FDI (International) system, Universal/ADA (US and Canada), and Palmer notation (UK, parts of Europe) all describe the same 32 permanent teeth and 20 primary teeth using completely different numbering schemes.

FDI #26 is Universal #14 is Palmer 6 (upper left first molar). The mapping is exact and non-controversial. But dental records frequently mix systems, abbreviate, or use shorthand that requires system-level context to parse correctly.

The knowledge base also addresses linguistic translation challenges specific to medical vocabulary. “Angina” in German medical usage typically refers to tonsillitis (angina tonsillaris), not chest pain. “Admission” in the UK NHS context carries administrative and clinical implications that differ from US hospital admission. These false cognates and terminology divergences sit above pure code mapping — and getting them wrong can have clinical consequences.

The Research Process

Building the knowledge base required structured research grounded in authoritative sources.

Primary sources form the foundation. These include national coding authority publications, official pharmaceutical databases, clinical coding standards from professional bodies, and published regulatory frameworks. We archive and timestamp these sources so that every production mapping can be traced back to its origin.

Every entry in the production knowledge base has been reviewed against authoritative sources. Approximate mappings are documented as approximate, with the reason for the gap noted. The knowledge base currently spans over 20 countries with thousands of validated code mappings.

Safety First

Medical translation carries patient safety implications that generic translation never does.

We build safety flags into the knowledge base at every level where clinical risk is present.

Drug Safety Alerts

Several high-frequency drugs in one healthcare system are banned, restricted, or have different risk profiles in another. The knowledge base includes drug-level safety flags that are surfaced in the translation output when these situations arise:

Metamizole (Novalgin, Dipyrone): Common OTC/prescription analgesic in Germany, Mexico, and India. Withdrawn from US market in 1977 due to agranulocytosis risk. A German prescription for Novalgin arriving at a US pharmacy needs an immediate alert, not just a name translation.
Nimesulide: NSAID available in Europe, India, and other markets. Not approved by the FDA. Associated with hepatotoxicity concerns that led to restrictions in multiple jurisdictions.
Codeine for pediatric use: Approved dosing and age restrictions vary significantly across healthcare systems. What is standard pediatric pain management in one system may be contraindicated in another.

Controlled Substance Differences

Scheduling categories vary by country. A drug that is Schedule IV in the US (moderate abuse potential, prescribable) may be an unscheduled OTC item in Germany, or vice versa. The knowledge base flags controlled substance classification differences so that the receiving clinician or pharmacist understands the regulatory context, not just the chemical identity.

OTC and Prescription Switches

Many drugs are prescription-only in the US and OTC in Europe, or the reverse. Ibuprofen at higher doses requires a prescription in Germany; lower-dose formulations are OTC. Several antihistamines that are OTC in the US are prescription-only in other markets. These are not pharmacologically trivial distinctions — they affect what patients can obtain, at what doses, and under what supervision.

Why Not AI-Only?

AI language models trained on medical text have impressive pattern-matching capabilities across medical codes. They can often produce plausible mappings for codes they’ve seen during training.

The problem is “often plausible” is not good enough for clinical use.

AI models hallucinate. They interpolate. They may produce confident, formatted-looking output for a code pair where no mapping exists — or where the mapping has important caveats that the model didn’t learn or didn’t surface.

More specifically: medical coding standards update annually. The ICD-10-GM is updated every January 1st by BfArM. ICD-10-CM updates take effect every October 1st. A language model trained on data through a certain cutoff date will be unaware of new codes, deprecated codes, and changed mappings that have been published since. A validated, timestamped knowledge base has known provenance. A language model does not.

We do use AI — for edge cases where no curated mapping exists, for reading clinical narrative, for generating translation summaries. But the curated knowledge base is the first, authoritative layer, and it outperforms AI-only approaches on every metric we’ve measured.

Continuous Improvement

Healthcare coding standards change every year. Regulatory decisions about drug approvals, withdrawals, and scheduling changes happen continuously. We update the knowledge base as these changes occur, with source-referenced updates so that the provenance chain is always maintained.

We also track translation cases where the knowledge base lacks coverage — these represent genuine gaps that get prioritized for the next research cycle.

How TranslateMD Uses the Knowledge Base

Every TranslateMD translation runs the source document’s detected codes against the knowledge base first. If a validated mapping exists, it’s used. If only an approximate mapping exists, the output is annotated with the confidence level and the nature of the gap. If no mapping exists, the case is routed to AI-augmented handling, and the output is clearly marked as AI-proposed.

The result is a translated document where the confidence level of every mapping is explicit — not a black-box output where the clinician has to guess which parts can be trusted and which parts need verification.

That transparency is the point. Medical translation is a tool for clinicians, not a replacement for clinical judgment. A well-annotated approximate mapping that prompts the right question is more useful than a confident wrong answer.

TranslateMD is available via API and web interface. For teams handling high-volume cross-border records, Team and Enterprise plans include full country coverage, bulk processing, and dedicated support.