Professor Patrik Eklund of Umeå University, Department of Computing Science, in Sweden, sent us an email with the title, “SNOMED's logical foundation is too shallow for practical applications”.
We attempted to set up a meeting with Prof. Eklund, Daniel and Mikael (both Sewdish), Jim (Case) and myself (not so Swedish). However, Prof Eklund declined to participate in a meeting involving DL experts.
Below we provide the substance of the email, and documents pertaining to Prof. Eklund’s field of study [https://confluence.ihtsdotools.org/display/mag/SNOMED%27s+logical+foundation+is+too+shallow+for+practical+applications].
We would welcome your responses to Prof. Eklund’s comment and what they mean, if anything, fos SNOMED, and IHTSDO.
“Description Logic as the underlying logic for health terminology is not
sufficient. I think it is not not sufficient at all. This means that
SNOMED is not as useful in practice as it could be. I have access to
SNOMED through the Swedish Socialstyrelsen (The National Board of Health
and Welfare) licence, and over the years I failed to see how SNOMED is
useful in practice. I have discussed these issues also with
Socialstyrelsen at several occasions.
Needless, when we come down to guidelines, and logical representation of
evidence in guidelines, the underlying logical framework must be rich
and precise. Statistics only as the underlying computational machinery
for "evidence" in EBM is not enough. Such "evidence" is not evidence in
the sense of logic.
Since SNOMED adopted DL, very little has happened regarding SNOMED's
underlying conceptual model. When IHTSDO adopted DL as the logic for
SNOMED I was very surprised. Throughout many years I have often said
that "ontology" in "health ontology" cannot be the same as "ontology" in
"web ontology". "Ontology" in "health ontology" is more, and more
specific to health classification. I find it very unfortunate that
IHTSDO complies with a bivalent and untyped logic like DL.
One problem is indeed DL's bivalence, and its restriction to using
unstructured relations. The same goes obviously for its sibling methods
like FCA or Rough Sets. SNOMED's concepts mostly hide underlying
multivalence, e.g. as needed when mapping to a terminological scope like
the ICF. ICF's generic scale is nicely multivalent, even if the 'not
specified' isn't logically interpreted. It can be, but WHO has not made
that observation.
Some say they have mapped SNOMED to ICF and vice versa, but these
mappings are very shallow constructions as they require to enforce ICF
into a DL framework using "is-a" as it is done in SNOMED. It is
basically just item mapping so that relations are preserved, and with
some ad hoc fine-tuning.
Bivalence can be easily extended to multivalence in SNOMED, but even
that is far from sufficient. The problem is still that "is-a" is a
relation basically over unstructured sets. It's also untyped in the
logical sense, and indeed, the underlying signature of DL is not a
signature (with sorts and operations) in the logical sense.
We have written several papers on this theoretical setting. Our
metalanguage is category theory, so our papers are probably a bit hard
to comprehend. However, we are developing practice based on our theory.
I am involved e.g. in information and process modelling within EIP AHA's
(European Innovation Partnership on Active and Healthy Ageing), so we
are continuously working with nomenclatures and classifications like WHO
classifications and potentially with SNOMED. EIP AHA's work on
assessment frameworks and upscaling is expected to benefit from these
logical considerations.”
2016-04-12 03:13
This is the first time I'm hearing about Prof. Eklund's work, but it seems to be addressing the issue of representing uncertainty in DL. There has been prior work in this area, e.g., PR-OWL: A Bayesian extension to the OWL Ontology Language (http://www.pr-owl.org/).
I'm not aware of any practical applications on a scale of SNOMED CT and description logic is hard even without taking into account fuzziness and uncertainty.
2016-04-12 08:17
I think there's really very little substance to this email. Specifically I can't see from this what "value" he's looking for; what specific problems does he want to try to solve using SNOMED CT and it's DL foundation?
Whether DL provides value to SNOMED CT depends on what value you're looking for. In Australia, I am firmly convinced that fully embracing the DL (via use of concrete domains) in the Australian Medicines Terminology delivered great value as it exposed mistakes in the modelling of the concepts. Thus the value derived was one of ensuring quality.
To construct a terminology on the scale of SNOMED CT without using something like DL but still retain the level of quality and extensibility would be very hard (yes, there are still problems, some that would be fixed by better/more use of DL - hello anatomy :-)
I think some of the issues raised around untyped sets are (partially) addressed extra-logically (outside the DL) by the MCRM, while others could be addressed with the introduction of concrete domains.
This is not to argue that SNOMED CT with DL is sufficient for all purposes. I think many people miss the point and expect that it and its foundations should be sufficient for more general knowledge representation and inferencing, rather than seeing it as a (reference) platform on which to build such things. I guess this is where he's going as he's looking for a semantics using truth values other than true & false. At this point in time, however, I am not aware of any obvious candidate semantics that one could select - one would need some compelling examples of value delivered for specific applications as well as evidence that such a semantics scales both from a computational perspective and from a human perspective; increasing the expressivity of a language makes it possible to say more complex things, but it also makes it easier to say (or imply) things you don't intend - look at how hard it is to model, at scale, with a DL that supports negation.
2016-04-12 09:15
I personally believe that this criticism is a mix of trivially true statements and some problematic interpretations. E.g. we all know DL is insufficient for modeling the clinical domain, and EL+ even more so, but we rely on other mechanisms to take some of the burden of the SNOMED CT logic, including MRCM but also e.g. information modeling frameworks for instance representation and guidelines/rules frameworks for uncertainty/probabilistic reasoning. The approach chosen is a divide-and-conquer approach. Still, if there are lessons to be learned from applying more expressive logic then let's look at the evidence and decide (thinking of Peter H's negation experiments, SemanticHealthNet's DL experiments, the EU DebugIT project applied a combination of DL and other FOPL subsets, and my own SNOMED CT scalability experiments). This concrete evidence is lacking in Patrik Eklund's argument.
BTW, Mikael and I are not Sewdish but Swedish
2016-04-12 07:14
I'm not the logician, but I wanted to have a crack at this paragraph, because I believe Prof. Eklund is missing the point about SNOMED CT.
Quote
Needless, when we come down to guidelines, and logical representation of
evidence in guidelines, the underlying logical framework must be rich
and precise. Statistics only as the underlying computational machinery
for "evidence" in EBM is not enough. Such "evidence" is not evidence in
the sense of logic.
If he is looking at logical representation of guidelines, then you must include the information model layer,
which we all know SNOMED CT is only part of the answer. I have recently been looking at some older work on GLIF,
GELLO and Virtual Medical record, which is one way to represent guidelines. There is a PDF on this from the Medical Objects team
in Australia here....
VMR/GELLO/GLIF