It was déjà vu all over again on day 2 of the AMIA 2008 Annual Symposium, when I attended an all-day tutorial titled Clinical Classifications and Biomedical Ontologies: Terminology Evolution, Principles and Practicalities (really — it was just plain déjà vu in my last post).
If a satellite or debris from a space shuttle exploding drops on your head, ICD-9 E845.9 is the code for you (“accident involving spacecraft injuring other person”). Luckily, this code’s never been needed, but unfortunately, E845.0, its sister concept, has been (“accident involving spacecraft injuring occupant of spacecraft”). In less extreme cases, we saw there were hundreds of codes for tuberculosis, but only one for prostate cancer (185), which is quite diverse in its forms and their reprercurssions. The cause is that whoever was in charge of the TB section of ICD-9 really “went to town”, with the implication that the urologists/oncologists were slackers.
Medical ontologies are being built the same way good-old-fashioned-AI ontologies were built back in the 1980s and early 1990s — with description logics.
I’m a big fan of description logics; I even wrote a book on the topic back when I was teaching description logics at CMU. To add to the sense of déjà vu, a former
victim student, Thomas Polzin of MultiModal, was in the small audience with me.
The uses to which description logics are being put in ontology development are in their sweet spot, which is complex type and consistency checking. It sure brought back memories to see a hierarchy of description logics on the complexity scale. It also reminded me that logics (by themselves) just aren’t good ways to program algorithms, which require a knowledge of compilation and execution, no matter how strong the siren call of the proponents of declarative this or that. That’s why I’ve always advocated using description logics for type checking in grammar development and logic programming (a position reinforced by Dick Crouch of Powerset and Tracy Holloway King of Xerox in their ACL software workshop talk this past summer).
The other thing that got me thinking about my more theoretical days was the emphasis on philosophy of language, which I also used to teach. The tutorial started at the “chasm of semantic despair” (somewhere between genomics and patient record data at around the cellular level, where terminology is unclear and evolving, though I think that’s just what the presenters had most experience with). We were pitched the “value proposition” (better organize your data if you want to improve quality of care; oddly not much talk of billing, which all goes through ICD-9 codes in the U.S. as of 2008). Then we started from the beginning with Aristotle. When talking ontologies, I’m more often reminded of Procrustes than of Aristotle.
Being scientists, the presenters were not surprisingly, although unbeknownst to them, squarely within the logical positivist tradition. Clearly they never read Wittgenstein or Quine, or they’d be post-analytic. Much less Rorty, or they’d be neo-pragmatic, like me. Keep tuned for more philosophy once I clear out the technical backlog.
The number of controlled vocabularies and ontologies in the medical domain is rather overwhelming. You’ve got ICD-9 (clinical modification of course, with ICD-10 scheduled for 2011 and ICD-11 in the works) for diseases and billing, Snomed 3 for clinical terminology, LOINC for lab terminology, NDC for drugs and packaging, and the UMLS to unify them all, with clinical additions like HL7 for electronic health records (Check out the U.S. National Cancer Institute’s BioPortal, which lets you search or browse some of these terminologies; we actually did some coding in the tutorial, which was illustrative.)
The tutorial was being taught by Chris Chute of Mayo Clinic and Jim Cimino of the NLM. As I learned later, this was like walking into an intro to Java tutorial being taught by Josh Bloch and Doug Lea. These guys are chairing the committees that are designing these ontologies (for instance, Chris Chute is the chair of the ICD-11 revision for the World Health Organization!).
This was one of the better tutorials I’ve ever seen, and I’d highly recommend it. It looks like they give this presentation regularly, judging by its practiced nature and their deck of 400 detailed PowerPoint slides. Unfortunately, being doctors, not computer scientists, I can’t find any of their slides or other reference materials on the web.