How the GALEN Intermediate Representation reconciles internal complexity

How the GALEN Intermediate Representation reconciles internal complexity with users’ requirements for appropriateness and simplicity A. Roberts*, W. D. Solomon*, J. E. Rogers*, C. J. Wroe*, A. L. Rector* *Department of Medical Informatics, University of Manchester, UK & Open. GALEN www. opengalen. org

Summary • The contradictory requirements on terminologies • Reconciling the requirements: a layered architecture • What does it give us? An example • Experiences of use • Conclusion

Terminologies need to be complex • Medicine is large - and growing • The requirements on terminologies are growing • EPR • Decision Support • Genotype meets Phenotype • New technologies are brought to bear on the problem • Description Logics

Users require simplicity severity mild cystitis moderate severe onset gradual sudden chronicity • Entering documentation into an EPR acute sub-acute chronic • Creating a clinical terminology

Reconciling complexity and simplicity: The GALEN Intermediate Representation GRAIL description logic The Open. GALEN model

GRAIL - a solid, formal ontology “Open fixation of a fracture of the neck of the left femur” (‘Surgical. Process’ which is. Mainly. Characterised. By (performance which is. Enactment. Of (‘Surgical. Fixing’ which has. Specific. Subprocess (‘Surgical. Accessing’ has. Surgical. Open. Closedness (Surgical. Open. Closedness which has. Absolute. State surgically. Open)) acts. Specifically. On (Pathological. Body. Structure which < involves Bone has. Unique. Associated. Process Fracturing. Process has. Specific. Location (Collum which is. Specific. Solid. Division. Of (Femur which has. Left. Right. Selector left. Selection))>))))

Intermediate Representation: a simpler layer for dissecting rubrics “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

Overview of the GALEN IR • Set of context sensitive operations • • • Mappings Filters Elisions Transformations Expansions • Related techniques • User interface perspectives • Language generation • Mapping to other representations

Overview of the GALEN IR dissection GALEN IR translator descriptors, links and sanctions descriptor and link mappings GRAIL expression Open. GALEN clinical terminology

What does the GALEN IR provide? • • Hides the complexity of GRAIL Semantic normalisation Abstracts specialised modelling knowledge A meeting point between traditional terminologies and GALEN

An example RUBRIC “Excision of lobe of lung” MAIN excising ACTS_ON lobe IS_PART_OF lung (Surgical. Deed which is. Mainly. Characterised. By (performance which. G is. Enactment. Of (Excising which plays. Clinical. Role Surgical. Role) which acts. Specifically. On (Lobe which is. Specific. Solid. DIvision. Of Lung)))

An example MAIN excisin ACTS_O g N lobe (Surgical. Deed which is. Mainly. Characterised. By (performance which. G is. Enactment. Of (Excising which plays. Clinical. Role Surgical. Role) which acts. Specifically. On (Lobe which is. Specific. Solid. DIvision. Of Lung))) IS_PART_ OF lung

An example MAIN excisin ACTS_O g N lobe IS_PART_ OF lung Surgical. Deed is. Mainly. Characterise d. By performance is. Enactment. Of Excising Plays. Clinical Role Surgical. Role acts. Specifically On Lobe is. Specific. Solid. Divisi on. Of Lung

Algorithm • Translate the start point • Recursively translate descriptor-link-descriptor triples • Descriptor to concept mapping • Apply descriptor mapping rules • Link to attribute mapping

An example (start point) MAIN excisin ACTS_O g N lobe IS_PART_ OF lung

Start point (start point) Surgical. Deed • Determines where the concept will be classified • Isolates and defers this classification decision • Separates knowledge of the underlying model from the knowledge required to dissect rubrics

An example (start point) Surgical. Deed MAIN excisin ACTS_O g N lobe IS_PART_ OF lung

Descriptor mapping excisin g Excising • Mapping to simple or compositional concepts • Semantic Normalisation via many to one mapping • Collapse synonyms onto a single concept • Cope with multiple source languages • Contextual mapping is possible • New descriptors may be added

An example (start point) Surgical. Deed MAIN excisin ACTS_O g N Excising lobe IS_PART_ OF lung

Mapping rules - specialisation Excising Plays. Clinical Role Surgical. Role If the descriptor maps to a kind of Generic. Process and follows one of a specified set of links, add the criterion “plays. Clinical. Role Surgical. Role” • Is it redundant? • Other rules - for example, concept wrapping

An example (start point) Surgical. Deed MAIN excisin ACTS_O g N Excising Plays. Clinical Role Surgical. Role lobe IS_PART_ OF lung

Link mapping MAIN is. Mainly. Characterise d. By performance is. Enactment. Of • Links map to GRAIL attribute-concept-attribute chains

An example (start point) MAIN Surgical. Deed is. Mainly. Characterise d. By performance is. Enactment. Of excisin ACTS_O g N lobe Excising Plays. Clinical Role Surgical. Role Lobe IS_PART_ OF lung

Link mapping ACTS_O N acts. Specifically On HAS_LOCATI ON • Provides semantic normalisation • Copes with variation in modelling style

An example (start point) MAIN Surgical. Deed is. Mainly. Characterise d. By performance is. Enactment. Of excisin ACTS_O g N lobe IS_PART_ OF lung Excising Plays. Clinical Role Surgical. Role acts. Specifically On Lobe Lung

Link mapping - finding candidates IS_PART_ OF is. Specific. Solid. Division Of is. Specific. Surface. Divisi on. Of etc • Link mapping is a two stage process • Initial mapping uses the descriptor context to disambiguate the link • This results in a set of possible GRAIL attributes

An example (start point) MAIN Surgical. Deed is. Mainly. Characterise d. By performance is. Enactment. Of excisin ACTS_O g N lobe IS_PART_ OF lung Excising Plays. Clinical Role Surgical. Role acts. Specifically On is. Specific. Solid. Division Of is. Specific. Surface. Divisi on. Of etc Lobe Lung

Link mapping - selecting a candidate is. Specific. Solid. Division Of is. Specific. Surface. Divisi on. Of etc is. Specific. Solid. Divisi on. Of • Final mapping chosen according to which attribute is sanctioned in the model • Hides the complexity of the GRAIL model • Isolates authors from changes in the underlying model - e. g. partonomic relationships

An example (start point) MAIN Surgical. Deed is. Mainly. Characterise d. By performance is. Enactment. Of excisin ACTS_O g N lobe IS_PART_ OF lung Excising Plays. Clinical Role Surgical. Role acts. Specifically On is. Specific. Solid. Division Of is. Specific. Surface. Divisi on. Of etc Lobe is. Specific. Solid. Divisi on. Of Lung

Experiences of use • GALEN-IN-USE • Trained 50 authors in 9 centers • Training takes 3 days (vs 3 months for GRAIL) • Dissecting averaged 50 rubrics person day • Drug ontology for Prodigy (prescribing guidelines) • Rapid modelling by a small number of authors • Initial knowledge capture • Semi-automated capture, semi-structured sources • Incomplete as a complete high level language

Update of ICPM-DE version 2. 4 Addition of rubrics in Musculo-Skeletal Surgery

Conclusion - the benefits of a layered architecture • Reduce the knowledge acquisition bottleneck • Speed, Training, Consistency • Separate clinical knowledge authoring from internal implementation • Accessible, decouple from implementation changes • Define simpler interfaces to a terminology • Semi-automated knowledge acquisition • A meeting point between GALEN and traditional terminologies

www. opengalen. org