Medication safety as a use case for argumentation mining We present a use case for argumentation mining, from biomedical informatics, specifically from medication safety. Tens of thousands of preventable medical errors occur in the U.S. each year, due to limitations in the information available to clinicians. Current knowledge sources about potential drug-drug interactions (PDDIs) often fail to provide essential management recommendations and differ significantly in their coverage, accuracy, and agreement. The Drug Interaction Knowledge Base Project (Boyce, 2006-present; dikb.org) is addressing this problem. Our current work is using knowledge representations and human annotation in order to represent clinically-relevant claims and evidence. Our data model incorporates an existing argumentation-focused ontology, the Micropublications Ontology. Further, to describe more specific information, such as the types of studies that allow inference of a particular type of claim, we are developing an evidence-focused ontology called DIDEO--Drug-drug Interaction and Drug-drug Interaction Evidence Ontology. On the curation side, we will describe how our research team is hand-extracting knowledge claims and evidence from the primary research literature, case reports, and FDA-approved drug labels for 65 drugs. We think that medication safety could be an important domain for applying automatic argumentation mining in the future. In discussions at Dagstuhl, we would like to investigate how current argumentation mining techniques might be used to scale up this work. We can also discuss possible implications for representing evidence from other biomedical domains. Talk for Dagstuhl Seminar 16161: Natural Language Argumentation: Mining, Processing, and Reasoning over Textual Arguments http://www.dagstuhl.de/en/program/calendar/semhp/?semnr=16161

1. Medication safety as a use case
for argumentation mining
Jodi Schneider and Richard D. Boyce
Dagstuhl Seminar 16161
Natural Language Argumentation: Mining, Processing, and Reasoning
over Textual Arguments
2016-04-19 1

2. Informatics
The management and processing of data, information
and knowledge.
2[Fourman 2002]

3. Informatics
The management and processing of data, information
and knowledge.
Examples:
o Biomedical informatics
o Dental informatics
o Legal informatics
o Business informatics
o Chemical informatics
o Neurinformatics
o ...
3[Fourman 2002]

4. Evidence Informatics
The management and processing of data, information
and knowledge ABOUT evidence.
4

5. Evidence Informatics
The management and processing of data, information
and knowledge ABOUT evidence.
Develop end-user applications.
e.g. Information retrieval using arguments & evidence.
o Kevin’s “legal argument roles”
o Benno’s PageRank for arguments
o Retrieve scientific articles by rhetorical or
argumentative features.
5

6. Evidence Informatics
The management and processing of data, information
and knowledge ABOUT evidence.
Develop end-user applications.
e.g. Information retrieval using arguments & evidence.
o Kevin’s “legal argument roles”
o Benno’s PageRank for arguments
o Retrieve scientific articles by rhetorical or
argumentative features.
Seek reusable underlying principles,
shared between several fields.
6

7. My approach to evidence informatics
o Understand user tasks and reasoning.
o Identify domain-specific argumentation schemes.
o Fill a knowledge base with arguments.
• Use domain-specific argumentation schemes as
templates.
• Fill “slots” in the scheme.
• Hand-annotate to bootstrap information extraction.
o Search engine for arguments and evidence
• Use rhetorical structures.
• Use argumentative structures.
7

8. MEDICATION SAFETY
8

9. Prescribers check for known drug interactions.
9

10. Prescribers consult drug compendia which are
maintained by expert pharmacists.
Medscape EpocratesMicromedex 2.0
10

11. Prescribers consult drug compendia which are
maintained by expert pharmacists.
Medscape EpocratesMicromedex 2.0
11

12. Problem
o Thousands of preventable medication errors occur
each year.
o Clinicians rely on information in drug compendia
(Physician’s Desk Reference, Medscape,
Micromedex, Epocrates, …).
o Compendia have information quality problems:
• differ significantly in their coverage, accuracy, and
agreement
• often fail to provide essential management
recommendations about prescription drugs
12

13. Problem
o Drug compendia synthesize drug interaction
evidence into knowledge claims but:
• Disagree on whether specific evidence items can support
or refute particular knowledge claims
13

14. Problem
o Drug compendia synthesize drug interaction
evidence into knowledge claims but:
• Disagree on whether specific evidence items can support
or refute particular knowledge claims
• May fail to include important evidence
14

15. “Addressing gaps in clinically useful
evidence on drug-drug interactions”
4-year project, U.S. National Library of Medicine R01
grant
(PI, Richard Boyce; R01LM011838)
o Evidence panel of domain experts: Carol Collins,
Amy Grizzle, Lisa Hines, John R Horn, Phil Empey,
Dan Malone
o Informaticists: Jodi Schneider, Harry Hochheiser,
Katrina Romagnoli, Samuel Rosko
o Ontologists: Mathias Brochhausen, Bill Hogan
o Programmers: Yifan Ning, Wen Zhang, Louisa
Zhang
15

16. Goals
o Long-term, provide drug compendia editors with
better information and better tools, to create the
information clinicians use.
o This talk focuses on how we might efficiently
acquire and represent
• knowledge claims about medication safety
• and their supporting evidence
o In a standard computable format.
16

17. DOMAIN-SPECIFIC
ARGUMENTATION
17

18. Drug Interaction Probability Score
1. Are there previous credible reports in humans?
• If there are case reports or prospective studies that clearly provide evidence
supporting the interaction, answer YES. For case reports, at least one case
should have a “possible” DIPS rating (score of 2 or higher).
• If a study appropriately designed to test for the interaction shows no
evidence of an interaction, answer NO.
…
5. Did the interaction remit upon de-challenge of the
precipitant drug with no change in the object drug? (if no
de-challenge, use Unknown or NA and skip Question 6)
• Stopping the precipitant drug should bring about resolution of the
interaction, even if the object drug is continued without change. …
• If dechallenge of the precipitant drug without a change in object drug did not
result in remission of the interaction, answer NO.
• If no dechallenge occurred, the doses of both drugs were altered, or no
information on dechallenge is provided, answer NA.
[Horn et al. 2007] 18

19. 19

20. [Hu et al. 2011] 20

21. 21[Hu et al. 2011]

22. 22[Hu et al. 2011]

23. 23[Hu et al. 2011]

24. [Boyce, DIKB, 2006-present] 24

25. [Boyce, DIKB, 2006-present] 25

26. DESIGNING AN EVIDENCE
BASE
26

27. Multiple layers of evidence
Medication Safety
Studies Layer
Clinical Studies and
Experiments
Scientific Evidence Layer
27

28. [Brochhausen, Schneider, Malone, Empey, Hogan and Boyce “Towards a foundational representation of
potential drug-drug interaction knowledge.” First International Workshop on Drug Interaction
Knowledge Representation (DIKR-2014) at ICBO.]
28

29. SCIENTIFIC EVIDENCE LAYER
29

30. Scientific Evidence Layer: Micropublications
[Clark, Ciccarese, Goble (2014) Micropublications: a semantic model for claims, evidence, arguments and
annotations in biomedical communications.]
30

31. Scientific Evidence Layer: Micropublications
[Clark, Ciccarese, Goble (2014) Micropublications: a semantic model for claims, evidence, arguments and
annotations in biomedical communications] 31

32. MODELING NARRATIVE
DOCUMENTS AS EVIDENCE
32

33. 33

34. 34

35. MP:Claim
35

36. 36

37. Building up an MP graph
37

38. Building up an MP graph
38

39. Building up an MP graph
39

40. Building up an MP graph
40

41. Building up an MP graph
41

42. Building up an MP graph
42

43. Building up an MP graph
43

44. Building up an MP graph
44

45. Building up an MP graph
45

46. HAND ANNOTATION TO
CREATE THE EVIDENCE BASE
46

47. Hand-extracting claims and evidence
o Sources
• Primary research literature
• Case reports
• FDA-approved drug labels
o Process
• Spreadsheets
• PDF annotation
47

48. 48

49. 49

50. Work to date
o 410 assertions and 519 evidence items transformed
from prior work.
o 609 evidence items (pharmacokinetic potential
drug-drug interactions) annotated by hand from 27
FDA-approved drug labels.
o 230 assertions of drug-drug interactions annotated
by hand from 158 non-regulatory documents,
including full text research articles.
50

51. DIRECTIONS & FUTURE WORK
51

52. We are developing a search/retrieval portal
It will:
o Integrate across multiple types of source materials
(FDA drug labels, scientific literature, …)
o Systematize search: Enable ALL drug compendium
editors to access the same info
o Provide direct access to source materials
• E.g. quotes in context
52

53. 53

54. Evaluation plan for the search/retrieval portal
o 20-person user study
o Measures of
• Completeness of information
• Level of agreement
• Time required
• Perceived ease of use
54

55. Generate multiple KBs from the same EB
55

56. My approach to evidence informatics
o Understand user tasks and reasoning
o Identify domain-specific argumentation schemes.
o Create arguments
• Use domain-specific argumentation schemes as
templates.
• Fill “slots” in the scheme.
• Hand-annotate to bootstrap information extraction.
• Automate.
o Provide argument and evidence-based information
retrieval
• Rhetorical functions
• Argumentative structures
56

57. Evidence modeling & curation
o Analogous processes could be used in other fields:
evidence modeling & curation is a general process.
o Biomedical curation is most mature: structured
nature of the evidence interpretation, existing
ontologies, trained curators, information extraction
and natural language processing pipelines
o Curation pipelines need to be designed with
stakeholders in mind.
57

58. Evidence Informatics
The management and processing of data, information
and knowledge ABOUT evidence.
Develop end-user applications such as using
arguments & evidence for information retrieval.
o Kevin’s “legal argument roles”
o Benno’s PageRank for arguments
o Retrieve scientific articles by rhetorical or
argumentative features
Seek reusable underlying principles,
shared between several fields.
58

59. Thanks to collaborators & funders
o Training grant T15LM007059 from the National
Library of Medicine and the National Institute of
Dental and Craniofacial Research
o The entire “Addressing gaps in clinically useful
evidence on drug-drug interactions” team from U.S.
National Library of Medicine R01 grant
(PI, Richard Boyce; R01LM011838) and other
collaborators
59

60. Jodi Schneider, Mathias Brochhausen, Samuel Rosko, Paolo Ciccarese,
William R. Hogan, Daniel Malone, Yifan Ning, Tim Clark and Richard D. Boyce.
“Formalizing knowledge and evidence about potential drug-drug
interactions.” International Workshop on Biomedical Data Mining, Modeling,
and Semantic Integration: A Promising Approach to Solving Unmet Medical
Needs (BDM2I 2015) at ISWC 2015 Bethlehem, Pennsylvania, USA.
Jodi Schneider, Paolo Ciccarese, Tim Clark and Richard D. Boyce. “Using the
Micropublications ontology and the Open Annotation Data Model to
represent evidence within a drug-drug interaction knowledge base.” 4th
Workshop on Linked Science 2014—Making Sense Out of Data (LISC2014) at
ISWC 2014 Riva de Garda, Italy.
Mathias Brochhausen, Jodi Schneider, Daniel Malone, Philip E. Empey, William
R. Hogan and Richard D. Boyce “Towards a foundational representation of
potential drug-drug interaction knowledge.” First International Workshop on
Drug Interaction Knowledge Representation (DIKR-2014) at the International
Conference on Biomedical Ontologies (ICBO 2014) Houston, Texas, USA.
Richard D. Boyce, John Horn, Oktie Hassanzadeh, Anita de Waard, Jodi
Schneider, Joanne S. Luciano, Majid Rastegar-Mojarad, Maria Liakata,
“Dynamic Enhancement of Drug Product Labels to Support Drug Safety,
Efficacy, and Effectiveness.” Journal of Biomedical Semantics. 4(5), 2013.
doi:10.1186/2041-1480-4-5 60

62. Medication Safety Studies Layer:
DIDEO
Brochhausen et al, work in progress, example of Clinical Trial
62

63. DIDEO: Drug-drug Interaction and Drug-
drug Interaction Evidence Ontology
63https://github.com/DIDEO

64. Definitions
o Drug-drug interaction
• A biological process that results in a clinically
meaningful change to the response of at least one co-
administrated drug.
o Potential drug-drug interaction
• POSSIBILITY of a drug-drug interaction
• Data from a clinical/physiological study OR reasonable
extrapolation about drug-drug interaction mechanisms
64

65. Other Implications
o Implications for ontology development.
o Implications for improving medication safety.
65

66. MEDICATION SAFETY DOMAIN
66

67. Existing approaches: Representation
Bradford-Hill criteria (1965)
1. Strength
2. Consistency
3. Specificity
4. Temporality
5. Biological gradient
6. Plausibility
7. Coherence
Bradford-Hill A. The Environment and Disease: Association or Causation?.
Proc R Soc Med. 1965;58:295-300.
67

68. Existing approaches: Representation
Horn, J. R., Hansten, P. D., & Chan, L. N. (2007). Proposal for a new tool to evaluate
drug interaction cases. Annals of Pharmacotherapy, 41(4), 674-680.
68

69. Existing approaches: Representation
Royal Dutch Association for the Advancement of
Pharmacy (2005)
1. Existence & quality of evidence on the interaction
2. Clinical relevance of the potential adverse reaction
resulting from the interaction
3. Risk factors identifying patient, medication or disease
characteristics for which the interaction is of special
importance
4. The incidence of the adverse reaction
Van Roon, E.N. et al: Clinical relevance of drug-drug interactions:
a structured assessment procedure. Drug Saf. 2005;28(12):1131-9.
69

70. Existing approaches: Representation
Boyce, DIKB, 2006-present 70

71. Existing approaches: Acquisition
o Evidence
71Boyce, DIKB, circa 2006

72. Silos: Multiple sources of information
Post-market studies
Reported in
Scientific literature
Pre-market studies Clinical experience
Drug product labels
(US Food and Drug
Administration)
72
Reported in

73. o What arguments are used in medication safety?
o How can these arguments be mined/identified?
o What work needs to be done?

74. Why is a new data model needed?
o Need computer integration
o Want a COMPUTABLE model that can make
inferences
74