Further Information on SNOMED-CT

In conjunction with his main article on The KIDMM Community's 'MetaKnowledge Mash-up, Conrad Taylor provides further information on SNOMED-CT.

SNOMED-CT is updated more frequantly than the alternatives, so responds quickly to changes in medical knowledge and practice. Thirty years ago, there was no classification for HIV/AIDS. Administrative terminology changes all the time; and individual specialists often ask for new terms to be added. Free text is valuable in medical records and there will always be a role for it, but it has two major drawbacks. Meaning may be ambiguous; and meaning is not available for computation. It cannot be analysed automatically for auditing and to direct payments by results; it cannot direct care pathways; and it cannot trigger automated warnings, such as about allergic reactions to penicillin or interactions between medications.

Concepts and Descriptions

Any SNOMED term is a class descriptor: 'jawbone' describes all jawbones. The boundary between a terminology and a classification is slim - it depends what you use it for. In linking to knowledge sources, one is not talking about this particular patient's jawbone fracture (as one would in a medical record), but a class of such fractures.

SNOMED-CT consists of hundreds of thousands of concepts, each given a unique 'ConceptID' code. A concept may have many names, each with a unique 'DescriptionID' code. ConceptID 22298006 has as its Fully Specified Name: Myocardial infarction (disorder) DescriptionID 751689013. Additionally, a preferred term is offered: Myocardial infarction - DescriptionID 37436014. There are also several synonyms such as Cardiac infarction or Heart attack, each with its own DescriptionID. (However, users of health information systems should never have to be confronted with these codes directly.)

This structure can resolve the ambiguities of free text. To a neurologist, 'cord compression' means compression of the spinal cord; to a midwife, the umbilical cord is implied. Within each context, the text term makes perfect sense Ð but as machines don't do context well, we need two distinct concepts in SNOMED-CT, with distinct ConceptIDs.

Does using a controlled terminology like SNOMED-CT limit what can be said in a medical record? Not much. With about 400,000 health care concepts, a million clinical terms and 1.5 million semantic relationships, SNOMEDÐCT provides a very long pick-list. Terms can also be qualified by contextual modification such as 'family history of' - 'planned' - 'refused'.

SNOMED-CT is organised as several hierarchies, based on multiple top-level concepts, an example of which is 'body structure'. Below that, index finger is 'a kind of' finger, which is 'a kind of' hand part. Most structure in each hierarchy is based on 'kind of' relationships. A single concept can live in more than one hierarchy: tuberculosis is a kind of 'disorder of the chest', but is also part of the 'infectious disease' hierarchy. In other words, SNOMED-CT is a 'polyhierarchy', not a taxonomy.

SNOMED-CT terms define an item, then qualify it. In making the definition, concepts can be combined with attribute-value pairs. For example a procedure can be represented thus:

method = excision
site = both tonsils
using = laser device

This is a post-coordinated representation of a clinical procedure. Its pre-coordinated equivalent would be 'bilateral laser tonsillectomy'. In the medical record, a practitioner may employ either the array of post-coordinated terms, or the pre-coordinated form. Most prefer to use pre-coordinated terms, within such a large terminology finding them is difficult: this is one of the tensions within SNOMED-CT.

Using SNOMED-CT in Practice

Finding the right term to use can be a challenge. You can search for a word, in which case you'll be presented with a list of all terms using that word. In a large terminology, that can be daunting! You can search for the pre-coordinated term, but there may not always be one available in SNOMED-CT. Or, you could start from a point you know, and browse the term hierarchies.These approaches can easily be combined.

The ideal situation for input is where the record-making system is designed for a well-defined area of practice. On a data input screen for recording blood pressure, the template can limit the choice of terms to only those that are relevant, on a drop-down menu.

Text could be automatically encoded as it is being entered. Such intelligent systems would be highly desirable, but at present they are unreliable. The codes they generate must be reviewed and approved before being committed to the medical record.

One problem in implementing SNOMED-CT is detecting equivalence between expressions of the same concept in pre-coordinated and post-coordinated representations. Most pre-coordinated terms in SNOMED-CT were taken in en masse from the Read Codes, and can represent quite complex conditions. An example is Colles' fracture - a pre-coordinated term for a fracture of the wrist end of the radius bone. Achieving canonical post-coordinated description of Colles' fracture involves several fundamental concepts, relationships between them, and qualifiers. Representing pre-coordinated identifiers canonically through relationships between fundamental concepts is not always yet available in SNOMED-CT.

This is significant: unless equivalence is detected 100%, all data retrieval has a grey margin the risk of possibly insufficient recall of all records to match the query. SNOMED-CT has problems expressing negation, which is critical in medical records - diabetes excluded, appendectomy not performed. The problem is, there are so many ways of expressing negation in the English language.

As the NHS starts to use SNOMED-CT to build, for example, data entry templates Ð errors are being discovered in it. Things are in the wrong hierarchies, or the wrong position. And there's a huge problem of enabling accurate, speedy use of SNOMED-CT in unconstrained situations, such as when taking a patient history. Until now, hospitals have not tackled the serious encoding of data about patients as it is collected; this will be the first time that many clinicians have met controlled terminology.

Human beings are lazy, and good at inference, so patient records are full of short cuts. 'BP 140/80' means 'blood pressure was taken; systolic pressure was observed to be 140 mm of mercury; diastolic pressure was 80 mm'. Any human reading this will assume it refers to the patient in whose record it is written, and that the reading was taken during a particular encounter. But computers are pernickety, and very bad at inference, and want everything made unambiguous.

Despite SNOMED-CT's great scope, it is not sufficient. It does not deal with numeric values, e.g. 'weight = 70 kg'. It does not identify individual objects in the world, such as people. Therefore SNOMED-CT terminology has to be used within an external syntax that binds instances of the SNOMED-CT concepts to their context, and this will include records of

  • who the record is about - the subject, typically a patient;
  • when and where the action/event occurred, or the observation was made;
  • who performed the action or made the observation.

Where Are We Now?

Since April 2007, an independent international body (IHTSDO) has been in charge of SNOMED-CT. Several countries have adopted it, with more on the way. There are no significant global rivals. It's been adopted by the NHS, but there's little practical experience of using it in patient record-keeping, and virtually no experience of using it in real time, during the patient encounter. It's a huge experiment - some may say, a huge gamble, with a lot at stake.

Return to main report

Return to top

Date published: 
Tuesday, 30 October 2007
Copyright statement: 

This article has been published under copyright; please see our access terms and copyright guidance regarding use of content from this article. See also our explanations of how to cite Ariadne articles for examples of bibliographic format.