Wednesday, July 15, 2009

Would Tightening the Definitions in HL7 Make Things Worse?

In a recent post I pointed to errors created by the flexibility of HL7, for example because it allows the NK code for 'next of kin' to be used in a message also, for example, when referring to a nurse. Graham Grieve has entered a Comment to this post, defending a view of the NK code as 'a field that users can put what they need into'.

As he points out:

The question is, for any particular use, is that use consistent with the definition of the field? I don't think it is in the hoot example. But consulting the standard, the definition is loose - like the definition of next of kin in the real world. Of course, in a competition to find the worst case of egragarious mis-use of a standard, HL7 regulars would line up to bring forward their favorite cases. But HL7 is not alone here. HTTP is the same too. As for web services ... . What this does betray is a flaw in the whole picture: the standards body defines a logical structure, but business use cases require things to be done, and how are they going to be done? So some partially educated analyst makes a (variably) bad decision under considerable pressure, usually with some input from actual implementers. The outcomes are unpredictable. Actually tightening the definitions (like in a proper ontology) can make things worse, and so can loosening them. Though either can make things better for at least some users in some circumstances. It's just something we have to live with.
I fear that Graham's relaxed attitude to the dilemmas of bad coding ignores an overwhelming problem in the current (v3) architecture of HL7, a problem which does much to explain the peculiar features of some of the discussions on HL7's email lists, and which explains also why HL7 needs to devote so much effort to the consideration of questions like 'Is Diet a subclass of SupplyAct?' or 'Is a car accident an intentional action?'

The problem is that the RIM rests on a counter-intuitive upper level ontology, one in which anything that is not an Entity has to be classified as an Act. This counter-intuitive ontology -- in which a disease, for want of any more appropriate category, is identified with an act of observation of a disease -- not only makes it hard for HL7 analysts to avoid making bad decisions; it is also likely to be responsible for medical (coding) errors downstream, and it also places obstacles in the way of simple remedies for such errors.

Consider, for example, the case described in part 4 of the interesting series "Are Health IT Designers, Testers and Purchasers Trying to Kill People?" by Scot M. Silverstein. The chart below represents a problem list presentation page generated in an Electronic Medical Record system via auto-population based on what its makers refer to as an 'artificial intelligence' function.


As Silverstein points out, a number of issues are illustrated here. For present purposes we focus only on those pertaining to the diagnosis 'atrial fibrillation.' This is, as Silverstein notes, an important piece of information for the clinician to know.

Except, however, when the patient does not have atrial fibrillation. This entry was auto-populated when a nurse ordered a blood clotting test and erroneously entered the reason for the test as "atrial fibrillation" (a common reason, just not the case here) to expedite the order's completion. Voila! Now Mrs. Jones carries this diagnosis, and the next clinician to come along might order her anticoagulated with heparin or coumadin for a history of A. fib, introducing yet more chance of an iatrogenic injury.

And I am told it takes going back to the vendor to have this erroneous entry permanently removed. Sheer idiocy! For instance, if the patient moves to a different unit, is discharged and returns to this hospital, or to an outpatient clinic or another hospital branch with this on the record, the chances of a screwup are not insignificant
HL7 was not, for all I know, involved in any way in the creation either of this system or of the specific problem list here illustrated. But consider the difficulties created by the RIM for any rectification. Suppose, for instance, one wished to have a facility to allow users of the EMR to generate different sorts of problem lists for different purposes. These might include, for example, problem lists that would include only those entries in which a disease is ascribed to a patient on the basis of some act of observation by a clinician. To achieve this end, we would need to keep records of observations by clinicians clearly separate from records of diseases on the side of the patient, along the lines illustrated for example here. We have attempted to identify a way in which the RIM could allow a programme do this coherently. But thus far we have not been successful.

1 comment:

InformaticsMD said...

Not sure what the problem you refer to is.

In EHR work I did a decade ago, the Dx table was in a one to many relationship with a table containing info on how the Dx was made (i.e., any number of "how made" descriptors could be added to each diagnosis). For instance, a Dx of "atrial fibrillation" might have been made by "EKG", "physical exam" and "auscultation."

(In the case you cite, the diagnosis was "made" by a mission hostile user interface that forced a user to enter it to get actual work done.)

S. Silverstein
MedInformaticsMD