ICD Code

When a doctor makes a diagnosis, she doesn’t just write it down as text. She also assigns it a short code made up of letters and digits. This code is called an ICD code. The abbreviation stands for “International Classification of Diseases,” that is, the international classification of diseases. This list is published by the World Health Organization, an agency of the United Nations. The idea behind it is simple: every disease should have the same unambiguous designation everywhere in the world.

Why diseases need a number

The same disease often has many names. Some doctors write “sugar sickness,” others “type 2 diabetes mellitus,” and still others abbreviate it. For people, this is usually understandable. For computers and for statistics, it’s a problem. The ICD code solves this: E11 always means type 2 diabetes, in Berlin just as in Tokyo.

In Germany, a great deal of money depends on this. Hospitals bill insurers for their services using so-called flat-rate case fees. Which flat rate is paid depends significantly on the coded diagnoses. An incorrectly assigned code can therefore mean that a clinic receives too little or too much money. An entire industry has grown up around this coding process, with specialized professionals and auditing bodies.

This also creates a huge data foundation. How many people get the flu? Is obesity increasing? Such questions can only be answered because diagnoses are recorded in a standardized, coded form. Cause-of-death statistics are also based on ICD codes.

Structure of the codes and the path to ICD-11

The codes are structured like a tree, from coarse to fine. The letter at the beginning stands for a large group of diseases. J, for example, denotes respiratory diseases. J45 is asthma, and additional digits after a decimal point allow the specific form to be pinpointed more precisely. This allows for a coarser or finer description as needed.

In Germany, the version currently in use is ICD-10-GM. GM stands for “German Modification,” i.e., a version adapted to the German healthcare system. It contains around 16,000 different codes and is revised every year. The World Health Organization has since published the successor version, ICD-11. This was built from the ground up for digital use, though the transition in Germany will still take years.

An important distinction: an ICD code describes a diagnosis, not a treatment. In Germany, operations and procedures are coded using a separate catalog called OPS. Only both systems together give the complete picture of a hospital case.

AI systems that code medical letters

Coding is tedious. Someone has to read long medical letters and pick out the appropriate codes from them. This is exactly where language models come in—AI programs that can understand and generate text. They read the documentation and suggest suitable codes. Experts call this “Computer-Assisted Coding.” A human confirms or corrects the suggestions, since errors in billing become costly.

For the AI industry, ICD codes are interesting for another reason as well. They make medical records machine-readable and thus analyzable. Research projects search through millions of coded cases, for instance, looking for patterns that point to risks. Without standardized codes, this would hardly be possible.

You may also encounter the ICD code quite directly. On a certificate of incapacity for work—the yellow slip from the doctor—the diagnosis is often given only as a code. Anyone who wants to know what’s behind it can look the code up in public directories. A common misconception here is that every code represents a confirmed diagnosis. There are additions for “suspected” or “status after,” which significantly change the meaning.