Our experience is that the data is the priority and so we want to focus on questionnaires to support capture of good quality data.
If you want to try to capture data from the majority of existing questionnaires then good luck – questionnaires notoriously ask questions badly, conflating multiple concepts into one question, Boolean True/False when there are other ‘shades of gray’ etc. They work variably as far as human interpretation but usually very badly wrt computer interpretation.
We do have experience in taking previous paper questionnaires, analysing the data requirements sought in terms of what we want to persist and then we design the UI/questions to match the data desired and/or suggesting the UI might show a questionnaire but each question the clinical data is actually recorded using core archetypes – for example “Do you have diabetes?” – ‘Yes’, is recorded using the value ‘Diabetes’ in the EVAL.problem_diagnosis and ‘No’ is recorded in the matching exclusion archetype. This creates real clinical data that can be used as part of a health record rather than create an electronic checkbox version of the original paper questionnaire which will never be used again, but capture dust in our EHR’s virtual archives.
In summary:
A generic question/answer pattern is next to useless - interoperability is really not helped, especially if both the question and answer has to be managed in the template. We have tried many variations of this in the past, some of which were uploaded into CKM and subsequently rejected.
Lock in those questionnaires that are ubiquitous, evidence based, validated as OBSERVATION archetypes and share them in the international CKM – eg AUDIT, Glasgow coma scale, Barthel index, Edinburgh post natal depression scale – there are many examples in CKM.
Lock in local questionnaires that are going to be reused in your organisation, region or jurisdiction even though they may not be reusable elsewhere. They will provide some interoperability even if might only be appropriate within one clinical system or national CKM. An example is the Modified Early Warning Score/National Early Warning Score – there are a few different variations used in different locations and whether they should all be in the international CKM is still not clear.
BTW Questionnaires should be modelled as OBSERVATIONs (ie evidence that can be collected over and over again using the same protocol) not EVALUATIONS (as they are not meta-analysis nor summaries).
A generic question/answer pattern is next to useless - interoperability is really not helped
I think you should rather say “A generic question/answer pattern is only useful for exchanging the questions and answers, and does not allow re-use of data”. This is not ‘next to useless for interoperability’, just not fit for any wider purpose
“• A generic question/answer pattern is next to useless - interoperability is really not helped, especially if both the question and answer has to be managed in the template.”
The complete sentence qualifies that the dependence on template modelling is the issue wrt interoperability. This is where a generic pattern is made specific for a given questionnaire or data set. Also that we have found there are multiple generic patterns, none of which is universally applicable and so to create multiple generic patterns becomes nonsensical.
In the templating scenario it is only if the exact same template is shared (where every question has been renamed and associated value sets inserted) that can we get any value. In our experience it is of higher value to create an archetype that can at least be shared locally and explicitly models the precise question/answer combo in order to achieve better reuse.
A few of generic ideas around the question-answer pair.
There are several kinds of question-answer pairs:
The general generic pattern is the pair: question - answer;
The questionnaire can be one or more question-answer pairs;
The questions can be locally defined or regionally, nationally, internationally used;
Answers can be free text, quantitative (number, code), semi-quantitative (derived from a categorised set of possible answers using inclusion and exclusion criteria) or qualitative (present, not present);
Answers can be aggregated by means of category, mathematical formula.
This list of kinds of questionnaires ranges from simple to very complex patterns.
Some are simple statements others are very complex scales and questionnaires in between.
They are all variations on a theme.
Any answer can expressed in the in-line local form and/or with a reference to an external source.
First of all, I largely agree with Heather re the current approach. At freshEHR, we generally try to maximise the use of international ‘semantic’ archetypes, including scales, scores etc but accept that this is often not necessary and that there is place for simply modelling aspects of the questionnaire as-is.
Commercially, I am interested in how we might make use of , or at worst, play nicely with the FHIR Questionnaire resources.
I read your post and your answer. Got two clear ideas:
generic approach / pattern is not clear or valuable.
current questionnaires might need a semantic review / corrections to check if a. questions are correctly asked, b. possible answers correspond to the question, c. the goal of the questions is clinical or has other goal, and that can determine if that question/answer should or not be part of the EHR
In my specific case, the scope might be narrower:
consider all questions are correctly modeled and answers correspond to questions
need the content to be archetyped and the data to be in the openEHR IM
most answer types will be boolean, coded text, text, and their null flavours
Since no generic approach might be possible, it might not be so bad to have a generic archetype for the questionnaire definition, just as a framework, and do the custom work on templates.
I thought about this for some time:
questions are elements with value alternatives for boolean, coded text, text
those elements have coded text names, that is where the question is specified, and codes can be custom (defined on templates)
of course, question occurrence is 1..*
on templates, specific occurrences are modeled, with specific codes for the element name, and those should have occurrence 1..1 (this should be possible in terms of the AOM/TOM but I doubt it is currently supported by modeling tools, I think TD doesn’t have this)
I know this is more an implementation idea, using standard modeling artifacts. But since there is no generic modeling for this, and the implementation needs to use the standard artifacts, I find this to be a not so bad solution.
In a generic questionnaire every data element will have to be an ‘Any’, defined in the template; the questions will need to be added, the relevant values defined etc.
AND the biggest problem for me is defining the pattern for the generic questionnaire – every time I’ve tried to design a flexible pattern that allows various levels of nesting under cluster headings for questionnaire groupings etc I get tied up in knots. There is no one single solution that will cover all questionairres – simple tree structures will not provide the solutions we need.
We could use a generic OBSERVATION container and a simple CLUSTER that allows multiple instances of a question with an ‘Any’ data type and a SLOT for nesting further instances. But the modelling overheads are onerous and I’m still struggling to see the value in pushing the work to the template design rather than just archetyping it. I’ve attached an example to show a possible pattern – but it is SO MESSY with renaming of every data point, constraining every aspect of each question (just as you would in a de novo archetype) and the overheads of explaining how to build a template from a generic pattern and define every single part of it in the template seem not worth the benefit.
questions and answers that depend on other input from previous question/answers or data in the database
and questionnaire answers that can be aggregated in one result.
Answers to questions sometimes will be queried and others are not.
Sometimes one question allows one answer, sometimes more than 1, sometimes questions are optional, sometimes not.
Sometimes these answers and questions play a role within the questionnaire, sometimes they are used on their own outside.
Sometimes the aggregated results will be queried.
Sometimes questionnaires are used locally for administrative reasons, sometimes they act like any other clinical test.
All these give rise to requirements for possible solutions.
‘Intelligent’ questionnaires I consider out of scope, for the moment.
And - I think- the problem is tractable on the condition that the questionnaire is modelled as CLUSTERs.
With one CLUSTER/pattern to allow the optional aggregated result and define the components of the questionnaire.
This organising CLUSTER/pattern makes use of CLUSTERs/patterns for each question/answer pair.
These CLUSTERs are topics that can be queried. One must allow that queries start at the CLUSTER level.
It must be possible to query the aggregated result and the individual topics.
