Strength Representation in Drug Model

The problem is we're not just dealing with more complexity for the ECL engine (of which there are already far too few that implement the full current feature set). These properties affect correct classification of the concepts which relies on the axioms, and there are other non-ECL implementations to consider too.

One approach would be to have the system store unit values at their lowest units (picogram, picoliter, etc.) and retain their original values at the presentation layer. Then results of differing units could be returned (e.g., mg and mcg).

To me the "presentation layer" in SNOMED CT are the descriptions - these are what is intended for human rendering. The underlying machinable layer are the axioms and the generated NNF that the ECL engines run off - these are intended for machine processing, i.e. the classifier for DL reasoning, ECL engines, and other implementations wanting to get at properties of the concepts. That data isn't designed for humans but for software engineers and machines.

We could create GCIs (ideally complex on both sides which we can't have at present) to provide unit conversions for all in use values/units to the classifier (these values could be calculated), and represent all relevant values and units in the calculated NNF. But that would result in quite a bit more weight for the terminology and classifier, and probably make the NNF look bloated, complex and confusing to those that don't know what they are looking at. You could keep all of those extra values and units out of the NNF and only use it in the ECL engines - but then what about other implementers who aren't ECL engines?

Our experience with implementers of medicines terminology content is that consistency and predicability in these properties is useful. For example it is much easier to sort a list of drugs by their ingredient strength if they are all predictably in the same units, than have to perform unit normalisation because you can't rely on the units being already consistent - and if the units are based on human preferences they won't be consistent. Anyone who's tried to write a name generator for medications will know that the rules for naming are complex and full of exceptions.

This is in part the problem with basing the DL definition of the concepts in "presentation strength" matching the way the drug's strength appears on its registration. The registration rules vary jurisdiction to jurisdiction, and even registration of different brands of essentially the same product within a jurisdiction. These strength representations also change over time - for example, a while back in Australia all the paracetamol liquids were in mg/5mL, now they are all mg/mL. Basing the definition in something arbitrary and temporal is problematic, when the products haven't changed but the way humans prefer to refer to and have them rendered does change.

To me this is more evidence that these unit preferences are the domain of preferred strength rendering and not defining. They have many different equivalent renderings, some of which may have been the preferred rendering in the past, or may become preferred in the future. The terminology can deal with these human rendering preferences in the part of the terminology designed for that - preferred terms - while the DL definition can remain stable over time.

The best way to make that definition stable is to define as simple as possible normalisation rules. These rules can ensure that classification is correct and consistent without the need for unit conversion chasing changing rendering preferences over time. It also provides implementation consistency and predictability in the NNF, without making it complex containing all the possible unit conversion values. It also makes ECL straightforward to write with the current ECL engines that don't need to be unit conversion aware.

Certainly if ECL engines or other implementations want or need to support unit conversions they could implement that algorithmically using UCUM for example to convert from user input to the consistent normalised units in the terminology (consistent target is important). Or they could be as you point out based on a precalculated index from the NNF with up and down scaled units...although this is theoretically unbounded. I think the algorithmic approach from the ECL input to a consistent and predictable substrate in the NNF is more tractable, particularly for arbitrary units input in ECL (i.e. maybe not matching any of the in use units in the NNF).