Medical Coding: Turning Words Into Counts

Part of Clinical Data Management and Integrity: Coding and Data Standards.

About 5 minutes. Free, in your language, no account needed.

Why this matters

A site records "felt sick in the morning." Another records "nausea." Another records "queasiness." A fourth writes "N+V" for nausea and vomiting.

Each is a faithful record of what a participant said. None can be counted with the others, and the study must report how many participants experienced nausea.

Medical coding maps what was said onto a controlled vocabulary so that the same clinical concept counts as the same thing regardless of how it was expressed. Every safety table in every regulatory submission rests on it.

The dictionaries

MedDRA, the Medical Dictionary for Regulatory Activities, codes medical conditions: adverse events, medical history, indications. It is hierarchical, with the verbatim term mapped to a preferred term, which sits within a high level term, a high level group term, and a system organ class.

The level at which results are presented changes what the reader sees. A table by preferred term shows "nausea" and "vomiting" as separate rows. A table by system organ class shows them together under gastrointestinal disorders. Neither is wrong, and choosing without stating the choice makes a safety profile look different for reasons that have nothing to do with the product.

WHODrug codes medications, mapping a reported product name to an active substance and its classification. This matters because the same substance is sold under many names in many countries, and a multinational study receives all of them.

Both dictionaries are versioned and both are updated regularly. Which version a study uses, and whether it re-codes when a new version is released, is a decision with consequences for comparability across studies in a program.

What coding is not

It is not a clerical task. Mapping "chest discomfort" requires deciding whether the reported experience is cardiac, musculoskeletal, gastrointestinal or anxiety-related, and the choice determines which safety signal, if any, the event contributes to. That is a clinical judgment and belongs to someone qualified to make it.

It is not a replacement for the verbatim term. The verbatim must be retained. It is the source, the coded term is a derived value, and the mapping must be reproducible from one to the other.

It is not neutral. Coding decisions shape the study's reported safety profile, and a systematic tendency to code toward less specific terms will dilute a genuine signal across several rows until it falls below the threshold at which anyone looks.

The autocoder and its limit

Most coding is automated against a dictionary of previously approved mappings, with unmatched terms going to a human.

That is efficient and it carries a specific risk: a mapping approved once is applied thereafter without review. An early wrong decision propagates silently across the study and, where the dictionary is shared, across the program.

The mitigation is periodic review of the approved mapping dictionary itself rather than only of the unmatched queue, and particular attention to the mappings applied most frequently, since those carry the most weight.

In practice

A data manager reviews a study's coding at the point where about a third of the events have been coded.

Two things stand out.

"Fatigue" and "tiredness" have been mapped to different preferred terms by two coders working at different times. Both mappings are defensible in isolation; together they split one clinical concept across two rows, and each row falls below the threshold at which the safety table reports it.

"Chest pain" has been mapped consistently to a musculoskeletal term following an early decision made for a participant whose chest pain was clearly musculoskeletal. Eleven subsequent participants with chest pain of unknown origin received the same mapping, because the autocoder applied the approved one.

Neither is a coding error in the usual sense. Both change what the safety table says, and the second could suppress a cardiac signal entirely.

What this looks like when it goes wrong

A study codes adverse events to a dictionary version current at study start. Enrollment runs four years and the dictionary is updated twice in that period.

The study does not re-code, which is a legitimate choice, and it does not document the choice.

At submission, the safety data is pooled with two other studies in the program that used later versions. A preferred term that was split into two more specific terms in the intervening versions now appears in one study as a single term and in the others as two.

The pooled analysis counts them as three different things. The apparent frequency of the original condition is lower than it is, distributed across terms that the pooling does not recognize as related.

The error is not in any study's coding. Each was internally consistent. It is in the absence of a decision, documented at the program level, about version alignment across studies intended to be pooled.

Key takeaways