Sharing healthcare
meaning
is hard to do (Why?)
John Madden
SNOMED/Duke University
A big, recurring mistake
…you’re
If you can’t
unlikely
specifytowhat
come
sortsup
of with
athings
good you’ll
be wanting
language
for
to say…
saying
it.
What do M.D.’s want to do/say?
• Store records
retrievably
• Ad hoc search
• Explore/mine records
• Support M.D.-machine
interaction (structured
reporting, templates)
• Decision support
• Artificial medical
intelligence
Okay
Oy vay
We constantly and unconsciously
hop heterogeneous levels of
abstraction !
Germ Theory of Disease
“Hmm, odd combination of symptoms!”
“Which facts belong in discharge plan?”
“Which facts belong in an eye exam note?”
“The Facts”
Why is this a problem?
• Specific, relatively tractable notion
of semantics underlies many very
successful knowledge
•representation
“Of what weformalisms
cannot speak, we
• must
“The existence
non-existence of
remain and
silent.”
atomic facts is Reality.…A proposition
presents the existence and nonexistence of atomic facts.”
But what’s missing?
•
•
•
•
Context
Relevance
Agendas
Workflow
embedding
• Paraconsistency
Three topologies for
“semantic interoperability”
Central semantic authority
What happens?
“If you would just
our controlled
vocabulary (damn
it), we wouldn’t
Spheres
of Influence
have all these
interoperability
issues.
Pro and con
• De facto
interoperabiltiy
• Inferencing is
tractable
• Maintenance is
conceptually
simple
• Terminological
complexity scales
poorly
• Inferencing is
brittle
• Restricts speech
• Maintenance (in
practice)
incredibly
laborious
Hierachical semantics
Devoid of specifics
but very sharable
Very specific but
poorly sharable
What happens?
• Centralized
terminologies ‘lose
weight”
• Users take on more
responsibility
• “Weakest link”
phenomenon
What happens?
Registries
Ontology
rconciliation
Pro and con
• Can still reason
• And maybe the
size of the fact
base is smaller
and more
manageable
• Unrecoverable
loss of
information
• Inconsistencies
still break
inferences
• And now they
might be harder
to pinpoint
Federated semantics
What happens?
• Mapping
• Responsibility is finely
divided
• Quality depends on
peer-to-peer
collaboration
• “Map-to-play”
• Few global guarantees
• Need to support a
“market” in ontology
fragments
Nirvana
• Locally consistent
• Globally tolerant of
inconsistency
• Finely granular in either case
At what level are we
non-interoperable?
Germ Theory of Disease
“Unusual combination of symptoms!”
“Which facts belong in discharge plan?”
“Which facts belong in my exam note?”
“The Facts”
Technology fit to use case
Very contextual
Heavy search
Heavy inferencing
Very non-contextual
Descargar

Document