1. Slide 1 IDO-WS 2010 Daniel Schober Daniel Schober on behalf of DebugIT Community Semantic integration of antibiotics resistance patterns

2. Slide 2 IDO-WS 2010 Daniel Schober Healthcare Context

3. Slide 3 IDO-WS 2010 Daniel Schober A need for ‚IT-biotics‘ DebugIT –Detecting and Eliminating Bacteria UsinG Information Technology –Using ‘semantic linked data’ to exploit distributed clinical data Acquire new knowledge –Through advanced data mining Apply knowledge in decision support –E.g. prescription choice Apply knowledge in monitoring –Analyze current & predict future trends –Discover patient safety patterns

4. Slide 4 IDO-WS 2010 Daniel Schober

5. Slide 5 IDO-WS 2010 Daniel Schober Using Ontologies in DebugIT Provide common semantic identifiers –Allow crosstalk within Interoperability Platform –SPARQL query to express research question –Provide formal meaning exploitable by logical & rule-based reasoners Integrate access to heterogeneous CIS –Normalization via terminologies and textmining

6. Slide 6 IDO-WS 2010 Daniel Schober Data normalisation & ontology mapping (annotation) Refined clinical data –uniform format & semantics –anonymized Ontologies Text Mining De-identification Raw clinical data –different encodings –different languages

7. Slide 7 IDO-WS 2010 Daniel Schober Data integration architecture ETL (2) populates local RDTBs in DMZ layer D2R conversion (3) allows SPARQL integration (4) via Ontologies (DCO, OO)

8. Slide 8 IDO-WS 2010 Daniel Schober Linking data values to ontologies (via CVs) 1.Textmining links CIS data values to CVs 2.Create SKOS mappings from CV to Ontology (DCO) SNOMED CT findings  Diseases Uniprot NEWT taxonomy  Bacteria WHO ATC codes  Drugs, antibiotics Foundational Model of Anatomy  Human anatomy …  …

9. Slide 9 IDO-WS 2010 Daniel Schober Ontology Layers within DebugIT 1 DebugIT Core Ontology (DCO) -Clinical domain of infectious diseases -OWL-DL 30 Operational ontologies (OO) -Implementation, module crosstalk, data mining -query building, statistics, analysis, evidences, maths, units, … -OWL-Full 7 Data Definition Ontologies (DDO) -Describing hospital specific CIS Data model

10. Slide 10 IDO-WS 2010 Daniel Schober Describing data Describing real world (independent of data) ‚female patient‘ in different ontology layers

11. Slide 11 IDO-WS 2010 Daniel Schober Steps for solving a clinical analysis question Clinician states clinical analysis question in natural language 1.Clinical Researcher –clinical analysis query via QueryBuilder & SPARQL OOs & DCO 2.Data Miners –data set queries for each targeted CDR via SPARQL DDOs 3.Data Manager –maintains N3 rule set to convert instances from the endpoint specific DDO to OO & DCO 4.Data Miners –aggregate data set SPARQL result graphs in DCO using the needed conversion rule sets –performs clinical analysis, e.g. using/creating N3 rules using OOs, DCO –formalizes the clinical analysis result, using OOs & DCO 5.Clinical Researcher –validates result & presents it to Clinician who validates result.

12. Slide 12 IDO-WS 2010 Daniel Schober Clinical Analysis SPARQL Query (construct) “What percentage of Escherichia coli cases, cultured from urine samples, is resistant to the combination of trimethoprim/sulfametoxazol (TMP/SMX) or trimethoprim in the period 2006-2010?” CONSTRUCT { ?percentage quex:percentageOf ?total; quex:percentageThat ?part; quex:hasValue ?percentageValue; quex:hasUnit units:percent. ?total rdfs:subClassOf cao:EColi, [ a owl:Restriction; owl:onProperty cao:culturedFrom; owl:someValuesFrom [ rdfs:subClassOf dco:UrineSample; a owl:Restriction; owl:onProperty biotop:outcomeOf; owl:someValuesFrom [ rdfs:subClassOf dco:UrineSampleCollection; a owl:Restriction; owl:onProperty event:during; owl:hasValue [ dco:hasStartDateTime "2006-01- 01T00:00:00"^^xsd:dataTime; dco:hasEndDateTime "2010-12- 31T23:59:59"^^xsd:dataTime]]]]. ?part rdfs:subClassOf ?total, [ a owl:Restriction; owl:onProperty cao:resistantTo; owl:someValuesFrom [ owl:unionOf (dco:Trimethoprim dco:SulfamethoxazoleAndTrimethoprim)]]}

13. Slide 13 IDO-WS 2010 Daniel Schober Clinical Analysis SPARQL Query (where) WHERE { ?percentage quex:percentageOf ?total; quex:percentageThat ?part; quex:hasValue ?percentageValue; quex:hasUnit units:percent. ?total rdfs:subClassOf cao:EColi, [ a owl:Restriction; owl:onProperty cao:culturedFrom; owl:someValuesFrom [ rdfs:subClassOf dco:UrineSample; a owl:Restriction; owl:onProperty biotop:outcomeOf; owl:someValuesFrom [ rdfs:subClassOf dco:UrineSampleCollection; a owl:Restriction; owl:onProperty event:during; owl:hasValue [ dco:hasStartDateTime "2006-01- 01T00:00:00"^^xsd:dataTime; dco:hasEndDateTime "2010-12- 31T23:59:59"^^xsd:dataTime]]]]. ?part rdfs:subClassOf ?total, [ a owl:Restriction; owl:onProperty cao:resistantTo; owl:someValuesFrom [ owl:unionOf (dco:Trimethoprim dco:SulfamethoxazoleAndTrimethoprim)]]}

14. Slide 14 IDO-WS 2010 Daniel Schober Data set SPARQL query (for HUG-DDO) CONSTRUCT { ?antibiogram a ddo:Antibiogram; ddo:hasCulture ?culturing; ddo:hasIdentifiedBacterium [ddo:hasBacteriumCode "562"^^biosko:uniProtTaxonomyDT]; ddo:hasTestedDrug [ddo:hasDrugCode ?atc]; ddo:hasOutcome ?antibiogramResult. ?culturing ddo:hasSampleType ?sampleType; ddo:hasResultDate ?resultDate} WHERE { ?antibiogram a ddo:Antibiogram; ddo:hasCulture ?culturing; ddo:hasIdentifiedBacterium [ddo:hasBacteriumCode "562"^^biosko:uniProtTaxonomyDT]; ddo:hasTestedDrug [ddo:hasDrugCode ?atc]; ddo:hasOutcome ?antibiogramResult. ?culturing ddo:hasSampleType ?sampleType; ddo:hasResultDate ?resultDate. FILTER (?atc = "J01EA01"^^clisko:atc20090101DT || ?atc = "J01EE01"^^clisko:atc20090101DT) FILTER ("2006-01-01T00:00:00"^^xsd:dateTime < ?resultDate && ?resultDate < "2010-12- 31T23:59:59"^^xsd:dateTime) FILTER (?sampleType = "102866000"^^clisko:sct20080731DT)} # to be changed to 122575003 for "Urine specimen"

15. Slide 15 IDO-WS 2010 Daniel Schober DDO to DCO mapping via N3 rules MAPPING FROM HUG-ddo:Culture TO dco:BacterialCultureProcedure {?culturing ddo:hasSampleType ?sample. ?Sample skos:exactMatch [skos:notation ?sample]} => {?culturing biotop:precededBy [a dco:SampleCollection; biotop:hasOutcome [a ?Sample]]}.

16. Slide 16 IDO-WS 2010 Daniel Schober Cross-site integrated SPARQL result 2 instances of total result set of 1764 a dco:AntimicrobialSusceptibilityTest, dco:BacterialAntibiogramAnalysis, dco:BacterialCultureProcedure; :hasOutcome [:encodes [:qualityLocated [a :SpeciesEscherichiaColiValueRegion]]], [ :encodes [:qualityLocated [a dco:Sensitive]]]; :hasParticipant [a dco:SulfamethoxazoleAndTrimethoprim]; dco:hasResultDateTime "2006-11-03T09:57:00"^^xsd:dateTime. a dco:AntimicrobialSusceptibilityTest, dco:BacterialAntibiogramAnalysis, dco:BacterialCultureProcedure; :hasOutcome [:encodes [:qualityLocated [a :SpeciesEscherichiaColiValueRegion]]], [ :encodes [:qualityLocated [a dco:Sensitive]]]; :hasParticipant [a dco:Trimethoprim ]; :precededBy [a dco:SampleCollection; :hasOutcome "abnormal urine" ]; dco:hasResultDateTime "2008-10-16T00:00:00"^^xsd:dateTime. …

17. Slide 17 IDO-WS 2010 Daniel Schober DCO design principles OWL-DL –Reasoner for autoclassification & consistency checks during OE –Reasoner infers multiple parenthood Reusing BioTop –Ensure a rigid modeling view –Provides reuseable constraints (bridges to all TLO) Concepts harvested from –Hospital CDR schemata –Competency questions from clinical use case Datadriven bottom up –Domain terminologies in use Via UMLS or OLS Ontology modularisation tools (A.Rector) HL7 v3 based

18. Slide 18 IDO-WS 2010 Daniel Schober DCO content (statistics) Ontology elements & axiomsOverallDCOBioTop Classes13111014375 Object Properties (relations)78374 Datatype Properties11100 Subclass Axioms14941050444 Equivalent Class Axioms1979899 Disjoint Axioms76175

19. Slide 19 IDO-WS 2010 Daniel Schober A tripartite granular disease model (SDP pattern)

20. Slide 20 IDO-WS 2010 Daniel Schober Inference of new facts (BloodSample is a BodyLiquidSample) Stated Facts Inferred Hierarchy (more structure) Logics Reasoner BodyLiquidSample = BloodSample = Asserted Hierarchy (flat list) BodyLiquid =

21. Slide 21 IDO-WS 2010 Daniel Schober Use CNL for Ontology Evaluation

22. Slide 22 IDO-WS 2010 Daniel Schober Next steps Enhance coverage Refinement of DCO structure –Addressing drugs dosages & disease therapies –Use rectors Snomed CT modularisation algorithm to extract relevant SNOMED CT IDs form DCO-provided seed list Publish and distribute –E.g. on Bioportal

23. Slide 23 IDO-WS 2010 Daniel Schober DCO evaluation Ultimate overall evaluation Can clinicians –run the overall system ? –build queries and understand results ? Can data miner –create data set results ? –do data mining and formalize quality criteria for results ? DCO internal evaluation Fitness for use tested by ability to answer CQ Evaluate validity of assertions by –Reasoners –Graphical and textual representations to domain experts –Serialization of modules into Constrained Natural Languages (CNL)

24. Slide 24 IDO-WS 2010 Daniel Schober (Preliminary) Conclusion Semantically rich application ontologies Successive Query formalisations are complex … but approach scales over space & time Used in practice –Practical SPARQL query building –Data integration across 7 EU Hospitals DL-reasoning helps ontology engineering –DL limitation justified for smaller ontologies –For larger models use rule-based reasoning As data is dirty we need we need to cope with errors arising

25. Slide 25 IDO-WS 2010 Daniel Schober Resources & Acknowledgements Resources DebugIT project –http://www.DebugIT.euhttp://www.DebugIT.eu Ontology sources –http://purl.org/imbi/dco/dcohttp://purl.org/imbi/dco/dco TermBrowser –http://www.imbi.uni-freiburg.de/~schober/dco_owlDoc/http://www.imbi.uni-freiburg.de/~schober/dco_owlDoc/ Acknowledgements Hans Cools, Martin Boeker, Kristof Depraetere, Douglas Teodoro, Remy Choquet, Stefan Schulz, Ilinca Tudose, Maren Kechel, Giovanni Mels, Dirk Coalert, Dimitris Iakovidis, the DebugIT team Funded by grant agreement ICT-2007.5.2-217139

26. Slide 26 IDO-WS 2010 Daniel Schober In the Hospital kitchen I was approached by a member of the feared ‘Antibiotics Resistance’ …