Talk given at the International Workshop on Knowledge Discovery from (Big) Text: Challenges and Opportunities when Mining Biomedical Text in Leuven on the 18th of May 2015
How to Troubleshoot Apps for the Modern Connected Worker
Medical Information Retrieval and its Evaluation: an Overview of CLEF eHealth Evaluation Task
1. Medical Information Retrieval and
its Evaluation: an Overview of CLEF
eHealth Evaluation Task
Lorraine Goeuriot
LIG – Université Grenoble Alpes (France)
lorraine.goeuriot@imag.fr
2. Presentation Overview
• Medical IR and its Evaluation
• CLEF eHealth
– Context and tasks
– IR tasks description
– Datasets
– Evaluation
– Participation
• Conclusion
2
3. Presentation Overview
• Medical IR and its Evaluation
• CLEF eHealth
– Context and tasks
– IR tasks description
– Datasets
– Evaluation
– Participation
• Conclusion
3
4. 4
Medical Professionals – Web
Search and Data
• Online information search on a regular basis
• Search failure for 2 patients out of 3
• PubMed search: very long (30+ minutes against 5
available)
• Knowledge production
constantly growing
• More and more publications
• Varying web access
6. 6
Patients and general public
• Change in the patient-physician relationship
• Patients more committed - cybercondria
• How can information quality be guaranteed?
10. Medical Information Retrieval
• How different is medical IR from general IR?
– Domain-specific search: narrowing down the
applications to improve results for categories of users
– Consequences of bad performances of a medical search
system
• Characteristics of medical IR:
– Data: medical/clinical reports, research papers, medical
websites…
– Information need: decision support,
technology/progress watch, education, daily care…
– Evaluation: relevance, readability, trustworthiness,
time
10
11. Evaluating Information Retrieval?
Did the user find the information she needed?
How many relevant documents did she get back?
What is a relevant document?
How many unrelevant document did she get back?
How long before she found the information?
Is she satisfied with the results?
…
Did the user find the information she needed?
How many relevant documents did she get back?
What is a relevant document?
How many unrelevant document did she get back?
How long before she found the information?
Is she satisfied with the results?
…
• Creation of (artificial) datasets representing a specific search
task, in order to compare various systems efficiency
• Involving human rating
• Shared with the community to improve IR
11
12. Typical IR Evaluation Dataset
Document Collection
Topic Set Relevance
Assessment
...
...
12
13. Existing Medical IR evaluation tasks
• Existing medical IR evaluation tasks:
TREC Medical Records 2011, 2012
TREC 2000 filtering track (corpus OHSUMED)
TREC genomics 2003-2007
ImageCLEFMed 2005-2013
TREC clinical decision support 2014, 2015
No patient-centered evaluation task
13
14. Presentation Overview
• Medical IR and its Evaluation
• CLEF eHealth
– Context and tasks
– IR tasks description
– Datasets
– Evaluation
– Participation
• Conclusion
14
15. CLEF eHealth
AP: 72 yo w/ ESRD on HD,
CAD, HTN, asthma, p/w
significant hyperkalemia &
associated arrythmias.
15
16. CLEF eHealth Tasks
2013
• Task 1: Named entity
recognition in clinical
text
• Task 2: acronym
normalization in clinical
text
• Task 3: User-centred
health IR
2014
• Task 1: Visual-Interactive
Search and Exploration
of eHealth Data
• Task 2: Information
extraction from clinical
text
• Task 3: User-centred
health IR
2015
• Task 1a: Clinical speech recognition from nurses handover
• Task 1b: Clinical named entity recognition in French
• Task 2: User-centred health IR 16
17. Presentation Overview
• Medical IR and its Evaluation
• CLEF eHealth
– Context and tasks
– IR tasks description
– Datasets
– Evaluation
– Participation
• Conclusion
17
19. IR Evaluation Task over the years
2013 2014 2015
Goal Help laypersons better
understand medical reports
Layperson checking
their symptoms
Topics 55 EN topics
built from
discharge
summaries
55 EN topics +
translation in
CZ, DE, FR
67 EN topics built from
images + translation in
AR, CZ, DE, FA, FR, IT,
PT
Documents Medical document collection provided by Khresmoi project
Relevance
assessment
Manual evaluation of relevance
of documents
Manual evaluation of
relevance and
readability of
documents
19
20. Presentation Overview
• Medical IR and its Evaluation
• CLEF eHealth
– Context and tasks
– IR tasks description
– Datasets
– Evaluation
– Participation
• Conclusion
20
21. Document Collection
• Web crawl of health-related documents (~ 1M)
• Made available through the Khresmoi project
(khresmoi.eu)
• Target: general public and medical professionals
• Broad range of medical topics covered
• Content:
• Health On the Net (HON) Foundation certified
websites (~60%)
• Various well-known medical websites: DrugBank,
Diagnosia, TRIP answers, etc. (~40%)
21
22. Topics &
context
Topics
2013
Manual creation
from randomly
selected annotation
of disorder in the
DS (context)
2014
Manual creation
from manually
identified main
disorders in the DS
(context)
2015
Manual creation from images describing
a medical problem (context)
22
23. Topics - Examples
<topic> <id>qtest3</id>
<discharge_summary>02115-010823-
DISCHARGE_SUMMARY.txt</discharge_summary>
<title>Asystolic arrest</title>
<desc>what is asystolic arrest</desc>
<narr>asystolic arrest and why does it cause death</narr>
<profile>A 87 year old woman with a stroke and asystolic arrest dies and
the daughter wants to know about asystolic arrest and what it
means.</profile>
</topic>
2013-2014
<topic> <id>clef2015.test.15</id>
<query>weird brown patches on skin</query>
</topic>
2015
23
25. Presentation Overview
• Medical IR and its Evaluation
• CLEF eHealth
– Context and tasks
– IR tasks description
– Datasets
– Evaluation
– Participation
• Conclusion
25
26. Guidelines for Submissions
26
Submission of up to 7 runs (per language):
Run 1 (mandatory) - team baseline: only title and
description fields, no external resources.
Runs 2-4 (optional) any experiment WITH the DS.
Runs 5-7 (optional) any experiment WITHOUT the DS.
2013 - 2014
Submission of up to 10 ranked runs (per language):
Run 1 (mandatory): baseline run
Runs 2-10: any experiment with any external resource
2015
27. Relevance Assessment
Manual relevance assessment conducted by medical
professionals and IR experts
4-point scale assessment mapped to a binary scale
– {0: non relevant, 1: on topic but unreliable} → non
relevant
– {2: somewhat relevant, 3: relevant} → relevant
4-point scale for NDCG and 2-point scale for precision
[2015] Manual assessment of the readability of the
documents conducted by the same assessors on a 4-
point scale
27
28. Relevance Assessment - Pools
Training set Test set
2013 Merged top 30 ranked
documents from Vector
Space Model and Okapi
BM25
Merged top 10 documents
from participants baseline
run, the highest two priority
runs with DS and highest
two without DS
2014
2015 Merged top 10 documents
from participants three
highest priority runs
28
30. Presentation Overview
• Medical IR and its Evaluation
• CLEF eHealth
– Context and tasks
– IR tasks description
– Datasets
– Evaluation
– Participation
• Conclusion
30
31. Participants and Runs
Monolingual IR Multilingual IR
# teams # runs # teams # runs
2013 9 48 -- --
2014 14 62 2 24
2015 12 92 1 35
31
36. Team-Mayo Team-AEHRCTeam-MEDINFO Team-UOG Team-THCIB Team-KC Team-UTHealth Team-QUT Team-OHSU
0
0.1
0.2
0.3
0.4
0.5
0.6
Baseline
Best run
36
2013 Participants' Results
Baseline vs best run
37. What Worked Well?
Team-Mayo:
• Markov Model Random Field to model query term
dependency
• QE using external collections
• Combination of indexing techniques + re-ranking
Team-AEHRC:
• Language Models with Dirichlet smoothing
• QE with spelling correction and acronym expansion
Team-MEDINFO: Query Likelihood Model
BM25 Baseline
37
39. What Worked Well?
Team-GRIUM:
• Hybrid IR approach (text-based and concept-based)`
• Language models
• Query expansion based on mutual information
Team-SNUMEDINFO:
• Language Models with Dirichlet smoothing
• QE with medical concepts
• Google translate
Team-KISTI:
• Language models
• Various QE approaches
39
41. 41
2013 - Use of Discharge Summaries
Team-Mayo Team-Medinfo Team-THCIB Team-KC Team-QUT
0
0.1
0.2
0.3
0.4
0.5
0.6
With DS
Without DS
Baseline
42. 42
How were DS used?
- Result re-ranking based on concepts extracted from
queries, relevant documents and DS (Team-Mayo)
- Query expansion:
* Filtering of non-relevant expansion terms/concepts
(Team-MEDINFO)
* Expansion with all concepts from query and DS (Team-
THCIB)
* Expansion with concepts identified in relevant passages
of the DS (Team-KC)
* Query refinement (Team-TOPSIG)
43. 2014 - Use of Discharge
Summaries
IRLabDAIICT KISTI NIJM
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
DS
No DS
43
44. How Were DS Used?
●Query expansion:
● Expansion using Metamap, with expansion
candidates filtered using the DS (Team-
SNUMEDINFO)
● Expansion with abbreviations and DS combined
with pseudo-relevance feedback (Team-KISTI)
● Expansion with MeSH terminology and DS (Team-
IRLABDAIICT)
● Expansion with terms from the DS (Team-
Nijmegen)
44
45. Presentation Overview
• Medical IR and its Evaluation
• CLEF eHealth
– Context and tasks
– IR tasks description
– Datasets
– Evaluation
– Participation
– Further analysis
• Conclusion
45
46. 46
Medical Queries Complexity
Query complexity = number of medical
concepts/entities it contains
radial neck fracture and healing time
facial cuts and scar tissue
nausea and vomiting and hematemesis
Dataset:
50 queries from CLEF eHealth 2013 (patients
queries)
Runs from 9 teams
Impact of the complexity on the systems
performances
48. Presentation Overview
• Medical IR and its Evaluation
• CLEF eHealth
– Context and tasks
– IR tasks description
– Datasets
– Evaluation
– Participation
– Further analysis
• Conclusion
48
49. Conclusion
• 3 successful years running CLEF eHealth
• Datasets are publicly available for research
purpose
• Used for research by organizers, participants,
and other groups
• Building a community – evaluation tasks,
workshop@SIGIR, special edition of JIR
49
50. For More Details
CLEF eHealth Lab overview:
Suominen et al. (2013). Overview of the ShARe/CLEF eHealth
Evaluation Lab 2013. In CLEF 2013 Proceedings.
Kelly et al. (2014). Overview of the ShARe/CLEF eHealth
Evaluation Lab 2014. In CLEF 2014 Proceedings.
CLEF eHealth IR task overview:
Goeuriot et al. (2013). ShAReCLEF eHealth Evaluation Lab
2013, Task 3: Information Retrieval to Address Patients’
Questions when Reading Clinical Reports. In CLEF 2013
Working notes.
Goeuriot et al. (2014). ShARe/CLEF eHealth Evaluation Lab
2014, Task 3: User-centred health information retrieval. In
CLEF 2013 Working notes.
50