Discussion 11 May 2021

Referred from 20202021-05-11 - SNOMED on FHIR Meeting (TB)

From Jeremy Rogers

The topic of flavours of Null cropped up here only last week, in the context of a draft new minimum dataset for radiotherapy procedures.  The designers want to capture where the radiotherapy beam was pointed at, and were proposing a permitted set of six values:

They were asking for a refset containing all six values, to be represented as SNOMED codes.  Our national architecture recommendation remains that such flavours of NULL must be captured only in the information model and should not be SNOMED coded.  But this means that the IM therefore carries all the burden for representing the processable semantics of what was unknown/unstated.  

Where clinically required, such IM representations would therefore ideally also support the capability to aggregate or abstract over such properties. So if somebody asks “I wonder how many records we’ve received from St Elsewhere’s recently where they claimed that the laterality of the intervention was ‘not known’?” ..then the information model must carry the burden of knowing all the flavours of elements across all the message information models that should count as a ‘lateralisation of an intervention’.

Our review of the wisdom of any possible design also touched on the strange behaviours of certain specific null flavours, such as “not asked”.  In some care settings, ‘not asked’ really means “not asked so not known by us. And since the patient has now left we will probably NEVER know the answer”. So it’s a permanent NULL.   But in other care settings – especially e.g. the answer to the question ‘preferred place of death’ in an End Of Life care pathway, the value “not asked” more commonly means ‘we considered it insensitive to ask so soon after diagnosis and so do not know the answer. But clinically you really need to ask them soon but when the time is right.”

We wondered how the different implicit workflows that come with the “not asked” value would be best handled, and whether or not that had any bearing on where it is best represented.  For example, do all existing clinical workflow authoring platforms enable clinicians to trigger clinical action alerts when such null data is passed, if the nature of the property that is ‘not asked’ is only represented in the information model construct containing that particular value?

From Ed Cheetham

Nice summary of the discussion. Probably implicit in your comments but worth repeating is the (more than theoretical) tension between a desire to name Null flavors explicitly (generally per ‘model attribute’ or ‘data item’) and a parallel desire to use SNOMED CT compositionally (and thus bury a proportion of those same attributes within expressions), presenting a challenge if trying to “…capture all Nulls in the information model…”.

In the context of negation, I see in the Condition resource that:

“It is common as part of checklists prior to admission, surgery, enrollment in trials, etc. to ask questions such as "are you pregnant", "do you have a history of hypertension", etc. This information should NOT be captured using the Condition resource but should instead be captured using QuestionnaireResponse or Observation. In this case, the combination of the question and answer would convey that a particular condition was not present.”

As far as I can see an almost perfect confusion between models of use and meaning!

Regarding (un)certainty – or more correctly one aspect of it – there’s a relatively recent piece of work here from a group that represent a combination of UK Royal Colleges. It’s particularly useful in that – whilst it name-checks a load of technical standards – I think its emphasis is strongly on clinical requirements and therefore un-tainted by our input. It also has a couple of tables (p34/5) giving some idea of how much machine-processable certainty is or can be used in primary and secondary care systems. ‘Diagnosis’ is also a good example of how fractal ‘uncertainty’ can be (“they’ve definitely got reflux but I’m not sure whether that’s the reason they’re ill” and “they’ve almost certainly had a stroke, probably caused by their confirmed AF’) and how solutions that apply uncertainty to statements ‘globally’ necessarily lose precision (in pursuit of attainability).