1. Making Terminologies useful and usable: Clinical Terminologies in the 21st Century: What are they for? What might they look like?
Alan Rector Bio and Health Informatics Forum/ Medical Informatics Group Department of Computer Science University of Manchester
img.man.ac.uk mygrid.man.ac.uk

2. An Old Problem
“On those remote pages it is written that animals are divided into:
a. those that belong to the Emperor
b. embalmed ones
c. those that are trained
d. suckling pigs
e. mermaids
f. fabulous ones
g. stray dogs
h. those that are included in this classification
i. those that tremble as if they were mad
j. innumerable ones
k. those drawn with a very fine camel's hair brush
l. others
m. those that have just broken a flower vase
n. those that resemble flies from a distance"
From The Celestial Emporium of Benevolent Knowledge, Borges

3. But why in healthcare? What’s it for? What’s the purpose?
Terminologies are of little use in themselves
How will it make care better? new things possible?
How will it make information systems better?
Painful experience of 20 years of over-selling and under performance
Do we need it: Clinically? Technically?
If we need it
what is ‘it’? Is ‘it’ one thing or many?
How will we know if we have ‘it’?
How will we know if ‘it’ is fit for purpose

4. Why Now? What’s different now?
Web, E-Science, Grids
Web speed
New technologies – OWL, new DLs, hybrid frame-DL environments
Post genomic medicine – personalised medicine
Joining up Healthcare Medical and Bioscience research – CLEF
Systemisation of healthcare
Clinical error reduction, clinical governance, evidence based medicine, …
Does anybody else have similar problems?
Ontologies are ‘flavour of the month’ in E-Science & Web
Bioinformatics is building them very rapidly
What can we learn from them?

5. Need more and better clinical information
A Convergence of Need
Post genomic research
Safe, high quality, evidence based health care
Need more and better clinical information
Which scales
In Size
In Complexity
Knowledge is Fractal

6. The requirements & Tools chain
Clinical users with needs to improve care / clinical knowledge
Applications for clinical users that meet those needs
Developers’ needs for terminology to build those applications
Terminologies which fit the applications’ builders’ needs to meet the clinical users’ needs

7. Who is it for? (Useful & usable to whom?)
Clinical users
Carers - prospective
Reviewers – retrospective
Researchers, managers, assessors, …
The community – how it shares its knowledge
Knowledge creators / distributors
Application developers
Easier to re-use what exists than to build new
Re-use or bust
Terminology authors
Quick responsive evolution

8. Useful and Usable Useful – for what? Usable – by whom?
Supports needed applications
Purpose
Does it well
Quality
Usable – by whom?
Intuitive / understandable
Handy
What you need is “to hand”
Timely

9. Preview of Arguments
The priorities are clinical needs supported by applications supported by terminology
Clinical quality is critical
Useful and usable to: clinical users, developers, ‘reviewers’, authors
In an open evolving world, open managed evolution is the only plausible way forward
Current technology gives us the opportunity to cope
Tools and environments are as important as content

10. Where we come from Best Practice Clinical Terminology
Data Entry PEN& PAD
Clinical Record
Decision Support
Best Practice
HealthCard
Mr Ivor Bigun
Any country
Anytown
Dun Roamin
GALEN Clinical Terminology
Data Entry Language Technology: CLEF
Electronic Health Records: CLEF
Decision Support & Aggregated Data
Best Practice

11. Terminology is Now Middleware human-machine / machine - machine
Explicit
Machines can only manipulate what is represented explicitly
More re-use  more manipulation  more explicitness
Understandable
People can only build, maintain and use it if they can understand it
Adequate
Expressive enough to do the job but still computationally tractable
Reliable
People can use it consistently
Scalable and maintainable

12. Where we think we are going:
Pre-1980: paper
Application specific retrospective human oriented systems
ICD, early SNOMED, CPT, OPCS, …
Mid 1980s – 1990s: “electronic paper”
Retrospective reporting + Prospective collection
ICPC Read I, II
Mid 1990s – mid 2000s:Centralised computer based
Retrospective reporting + Prospective collection
OpenGALEN, Read III, SNOMED-RT…
PEN&PAD
Mid 2000s – ?: Web based open managed evolution
???? – but see the Semantic Web, Gene Ontology, etc.

13. How we will know when we get there Criteria for success
Re-use
A recognised growing library of common decsision support modules
Stop starting from scratch!
Integration
2+ independently developed DSSs integrated with 2+ independently developed EPRS without exponentially increasing effort.

14. Criteria for success Authoring Indexing
No individual invests in their own terminology
enterprise-wide terminology servers
Indexing
Simplification of systems
a sharp drop in special cases and exceptions
a sharp increase in authors’ productivity

15. Criteria for success User interfaces
Real systems in real use with real patients by real clinicians
transparent systems

16. Stones in the Road Why are we not there yet?
Some background definitions
Some hypotheses

17. Clinical quality & logical quality
Clinical quality – do users put in the right things?
Repeatability of information captue (inter rater reliability)
For decision support in prospective use
For retrieval in retrospective use
Salience
Relevance to clinical decisions for prospective use
Significance to questions for retrospective use
A better measue than “coverage”
Logical quality – do systems give the right responses?
Correct organisation (classification)
Correct inferences given correct input

18. Hypothesis 1
Most computer oriented terminology development ignores clinical quality …
The EHR as black hole
Bigger is not necessarily better
…although clinical quality was the primary concern of traditional paper/human oriented terminologies (and there are honourable exceptions – e,g, ICPC).
Evidence: High variability in recorded use Systematic failure to use data from GP systems in clinical studies (despite PRIMIS) Our own & colleagues’ experience in repeated studies Current planned cost of cohort ‘post genomic’ studies

19. Three models Meaning - ontologies
Can I depend on the answers?
“Dyspnoea is a respiratory problem”
Clinical significance – decision support
What should I think of / how does it affect decisions
“Dyspnoea can be a symptom of congestive heart failure”
Model of use – EHR/human factors
Is what I want ‘to hand’ – is it ‘handy’?”
“Dyspnoea should be a question on a cardiac history”

20. Hypothesis 2
Early terminologies emphasised models of use and significance and failed for lack of model of meaning
“Heart diseases” are in 13 Chapters of ICD9
Recent terminologies emphasise model of meaning and fail for lack of models of use and significance
Evidence:
User dissatisfaction, non-use, and poor quality data
The few systems based on models of use have been surprisingly popular with doctors, e.g. MedCin, ORCA
But hard to use for retrieval
We have fewer formal models of use than of meaning
We have almost no models of ‘significance’

21. Grounding cost vs Clean-up cost (with thanks to Enrico Coiera)
The cost of establishing a given quality of communication
How much French do you need to order a meal?
“Clean up cost”
The cost of fixing miscommunication
How many surprises will you accept? of what kind?

22. Special purpose vs Re-usable Multipurpose
Special purpose terminologies
Almost all retrospective
Reporting for remuneration – ICD9-CM, CPT
Reporting for epidemiology - ICD10, OPCS
Multipurpose re-usable terminologies
Aspire to be the glue for ‘Patient centred systems’ & ‘Personalised Medicine’
Decision support
Electronic Health Records
Research
Integration with Bioscience …
But too often ‘multipurpose’ means ‘no purpose’ ‘multiapplication’ means ‘no application’

23. Need “Multipurpose” mean “no purpose”?
Multiple purposes held by multiple groups
Multiple sources of expertise & authority
One size does not fit all
Multiple collaborations
Multiple legacies
Multiple purposes use multiple applications
Applications are the point of interaction
Applications make needs concrete & testable

24. Multipurpose means interacting with others It’s a big open world out there…
Bioscience
Gene Ontology, National Cancer Institute Center for Bioinformatics (NCICB), The Digital Anatomist/ Mouse Anatomy/Mammalian Anatomy, BioJava, PRINTS, EMBL, Microarrays, Protemoics, Metabalomics, Systems Biology…
Medicine meets bioscience
Cancer therapeutics, New imaging, …
E-Health: sharing and pooling data: Collections based research”
BioBank, NTRAC, NCRI, NCTR, CLEF, …
“Health Intelligence”
MRC policy on data sharing …

25. Hypothesis 3
Grounding costs can be delimited for special purpose terminologies
Grounding costs are indefinite for re- usable terminologies (& is historically high)
Without purposes testable through applications there
Danger of the escalating deadly embrace
“Must have terminology to build applications; but Must have applications before terminology”
Evolutionary approach the only exit

26. Central Control vs Open managed evolution
Académie française vs Oxford English Dictionary
Scholasticism vs Empiricism
The ‘arrogance of the a prior’ People don’t know what they do
Look to see what is actually used
Language technology shows time and again that our predictions are faulty
Command economy vs Social Market
Participation is the issue rather than money
Somebody will still have to pay
But at least they might pay for something useful

27. Central management Owned by one “Authority”
Coupling tight / autonomy low/ participation low
“Grounding costs” high / “Clean up costs” low?
must have everything before you can do anything
Change slow & lockstep
A product

28. Open managed evolution
“Owned” by the community – multiple “authorities”
Coupling loose/ autonomy high / participation high
To be useful & usable involve users using systems
“Grounding costs” low / “Clean up costs” high?
“Just in time” “Just enough”
Agree where it counts
Change quick and local - “threaded with annealing”
A process

29. Hypotheses 4
Single purpose clinical terminologies can be best managed centrally
By definition are developed in conjunction with an application
Re-usable terminologies can only succeed by open managed evolution
Many purposes require many contributors
Evidence: Speed of uptake of HL7/LOINC
W3C & the evolution of the Web
Re-usable terminologies can only be developed in open collaboration with applications
Otherwise “multipurpose” become “no purpose”

30. Hypothesis 5
Modern technology provides the means to support open managed evolution without compromising clinical quality or technical stability
Trade lower grounding cost for greater clean up cost
Focus on minimal stable core. Defer commitments.
Evidence: OpenGALEN, Gene Ontology
Utilise Web/Grid technologies for rapid dissemination and coordination
Evidence: Current developments at Mayo clinic using LDAP
Distribute terminology like domain names

31. The technologies
Applications centric development Decoupled development
Special purpose languages / “Intermediate Representations”
Deferred commitment
Clinical before technical
Logic based ontologies +
Models of clinical significance
Models of clinical use
Models of EHRs
Web services & Grid technology
Authentication/authorisation/accounting
Distributed directories & LDAP
Service discovery

32. Decoupled development using “Conceptual Lego”
If we manage the connectors and the pieces the users can build most things for themselves
Without compromising quality

33. Applications centric Development
Meta-authoring
Common Terminology/ Ontology
clinical
applications
authoring
environments
Intermediate Representations
clinicians / Applications builders Empowered Authors
templates/ views

34. Loosely Coupled Development
Local author uses resources & templates to formulate definition
templates
Worldwide Resources
problems
Server validates & organises
Central Ontology
Central Gurus integrate & fix problems
Local Author needs new terms for application
Local author checks
Local Ontology
updates

35. The templates are more important than the underlying formalism…
"Open fixation of a fracture of the neck of the left femur"
MAIN fixing
ACTS_ON fracture
HAS_LOCATION neck of long bone
IS_PART_OF femur
HAS_LATERALITY left
HAS_APPROACH open
“Intermediate Representations” are critical

36. …complex underpinnings can &will change
(‘SurgicalProcess’ which
isMainlyCharacterisedBy (performance which
isEnactmentOf (‘SurgicalFixing’ which
hasSpecificSubprocess (‘SurgicalAccessing’
hasSurgicalOpenClosedness (SurgicalOpenClosedness which
hasAbsoluteState surgicallyOpen))
actsSpecificallyOn (PathologicalBodyStructure which <
involves Bone
hasUniqueAssociatedProcess FracturingProcess
hasSpecificLocation (Collum which
isSpecificSolidDivisionOf (Femur which
hasLeftRightSelector leftSelection))>))))

37. Decoupling & Flexibility
Use formality to permit flexibility
Change need not mean instability
Formality means effects can be predited
Most users only need change in tightly controlled areas
Lesson from the Semantic Web: “Forking” a natural part of development
Harmless if strictly local
Manageable if controlled from standard “Lego” & templates
“Clean up cost”
10%-20% central effort is a reasonable target
Necessary to cope with change and ignorance
Evolution by “annealing”

38. Scalable models of use

39. Scalable models of Use: PEN&PAD
Structured Data Entry
File Edit Help
FRACTURE SURGERY
Reduction
Fixation
Fixation
Open
Closed
Open
Tibia
Fibula
Ankle
More...
Radius
Ulna
Wrist
Humerus
Femur
Femur
Left
Left
Right
More...
Gt Troch
Shaft
Neck
Neck
250,000 forms from 10,000 Facts “Fractal tailoring”

40. Scalable models of use: Fractal tailoring forms for clinical trials
Hypertension
Hypertension
Idiopathic Hypertension
Idiopathic Hypertension`
In our company’s studies
In our company’s studies
Idiopathic Hypertension in Study a phase 2
Idiopathic Hypertension in our co’s phase 2 study a
In Phase 2 studies
In Phase 2 studies

41. It can work The Lessons of GALEN The Lessons of PEN&PAD
Loosely coupled development based on formal ontologies works
“Coherence without uniformity”
90% of work done locally
Ontologies can be modular rather than monolithic
“Plug and play” terminology development
The Lessons of PEN&PAD
Models of use based on formal ontologies scale
250,000+ forms from 10,000 ‘facts’
The Lessons of the Semantic Web
It works for knowledge management
Growing user community outside of medicine
No longer “rocket science”

42. So what are “Logic based ontologies”

43. Logic-based Ontologies: Conceptual Lego
“SNPolymorphism of CFTRGene causing Defect in MembraneTransport of ChlorideIon causing Increase in Viscosity of Mucus in CysticFibrosis…”
“Hand which is anatomically normal”

44. Logic based ontologies
A formalisation of semantic nets, frame systems, and object hierarchies via KL-ONE and KRL
“is-kind-of” = “implies” (“logical subsumption”)
“Dog is a kind of wolf” means “All dogs are wolves”
Modern examples: DAML+OIL /“OWL”?)
Older variants LOOM, CLASSIC, BACK, GRAIL, K-REP, …

45. Logic Based Ontologies: The basics
Validating
(constraining cross products)
Primitives
Descriptions
Definitions
Reasoning
Thing
Feature
pathological
red
Structure
Encrustation
+ involves: MitralValve
Thing
+ feature: pathological
Structure
+ involves: Heart
Heart
MitralValve
Encrustation
MitralValve
* ALWAYS partOf: Heart
Encrustation
* ALWAYS feature: pathological
red
+ partOf: Heart
red
+ partOf: Heart
+ (feature: pathological)

46. Building with Conceptual Lego
Species
Genes
CFTRGene in humans
Protein
Protein coded by (CFTRgene & in humans)
Function
Membrane transport mediated by (Protein coded by (CFTRgene in humans))
Disease
Disease caused by (abnormality in (Membrane transport mediated by (Protein coded by (CTFR gene & in humans))))

47. Avoiding combinatorial explosions
The “Exploding Bicycle” From “phrase book” to “dictionary + grammar”
ICD-9 (E826) 8
READ-2 (T30..) 81
READ-3 87
ICD-10 (V10-19 Australian) 587
V31.22 Occupant of three-wheeled motor vehicle injured in collision with pedal cycle, person on outside of vehicle, nontraffic accident, while working for income
and meanwhile elsewhere in ICD-10
W65.40 Drowning and submersion while in bath-tub, street and highway, while engaged in sports activity
X35.44 Victim of volcanic eruption, street and highway, while resting, sleeping, eating or engaging in other vital activities

48. The Cost: Normalising (untangling) Ontologies
Structure
Function
Part-whole
Structure
Function
Part-whole

49. The Cost: Normalising (untangling) Ontologies Making each meaning explicit and separate
PhysSubstance Protein ‘ ProteinHormone’ Insulin ‘Enzyme’
Steroid ‘SteroidHormone’
‘Hormone’
‘ProteinHormone’ Insulin^ ‘SteroidHormone’ ‘Catalyst’ ‘Enzyme’
PhysSubstance Protein ProteinHormone Insulin Enzyme Steroid SteroidHormone
Hormone
ProteinHormone^ Insulin^ SteroidHormone^ Catalyst Enzyme^
…build it all by combining simple trees
… ActionRole PhysiologicRole HormoneRole CatalystRole …
… Substance BodySubstance Protein Insulin Steroid …
Hormone = Substance & playsRole-HormoneRole
ProteinHormone = Protein & playsRole-HormoneRole SteroidHormone = Steroid & playsRole-HormoneRole
Catalyst = Substance & playsRole CatalystRole
Insulin  playsRole HormoneRole
Enzyme ?=? Protein & playsRole-CatalystRole

50. But none of it works without tools None of it works without communication & cooperation

51. Communicating software environments “Environments” rather than “servers”
Clinical users - care and review
Environments for entering& retrieving information
Methodologies for measuring and monitoring quality of information
Human factors, language technology, fractal tailoring to needs
Application developers
Configuration tools – much more than “terminology servers”
The key to success
Ontology authors
Tools for distributed loosely coupled authoring
Ontology managers (the “gurus”)
Tools for reconciliation, change management, & meta-authoring of templates

52. Summary of Arguments
The priorities are clinical needs supported by applications supported by terminology
Unless they serve clinical needs, applications are useless
Unless they serve applications, terminologies are useless
Unless used reliably, terminologies are meaningless
“Meaning is a social construct”
Clinical quality should be our watchword
Useful and usable to: clinical users, developers, ‘reviewers’, authors
Requires models of use & clinical significance
Requires tools and environments
In an open evolving world, open managed evolution is the only plausible way forward
Participation and control are the issues – not money
Current technology gives us the opportunity to cope
If we let development follow need
If we use them to the full
19th century methods won’t cope with 21st century problems

53. Making Terminologies useful and usable: Clinical Terminologies in the 21st Century: What are they for? What might they look like?
Alan Rector Bio and Health Informatics Forum/ Medical Informatics Group Department of Computer Science University of Manchester
img.man.ac.uk mygrid.man.ac.uk