This Provisional PDF corresponds to the article as it appeared upon acceptance. The fully-formatted
© 2005 Fontelo et al., licene BioMed Central Ltd.
This is an Open Access article distributed under the terms of the Creative Commons Attribution Licen (creativecommons/licens/by/2.0), which permits unrestricted u, distribution, and reproduction in any medium, provided the original work is properly cited.
1
ask MEDLINE: a free-text, natural language query tool for MEDLINE/PubMed
Paul Fontelo*§, Fang Liu*, Michael Ackerman*
Office of High Performance Computing and Communications, National Library of Medicine, 8600 Rockville Pike, Bethesda, Maryland 20894, USA.
*The authors contributed equally to this work
§Corresponding author
去黄
Email address:
PF: fontelo@v
FL: fliu@v
MA: ackerman@v
2
Abstract
Background
Plain language arch tools for MEDLINE/PubMed are few. We wanted to develop a arch tool that would allow anyone using a free-text, natural language query and without knowing specialized vocabularies that an expert archer might u, to find relevant citations in MEDLINE/PubMed. This tool would translate a question into an efficient arch.副翼
Results
The accuracy and relevance of retrieved citations were compared to references cited in
BMJ POEMs and CATs (critically appraid topics) questions from the University of Michigan Department of Pediatrics. askMEDLINE correctly matched the cited references 75.8% in POEMs and 89.2 % in CATs questions on first pass. When articles that were deemed to be relevant to the clinical questions were included, the overall efficiency in retrieving journal articles was 96.8% (POEMs) and 96.3% (CATs.)
Conclusions
ask MEDLINE might be a uful arch tool for clinicians, rearchers, and other information ekers interested in finding current evidence in MEDLINE/PubMed. The
text-only format could be convenient for urs with wireless handheld devices and tho with low-bandwidth connections in remote locations.
Background
ask MEDLINE (askmedline.v) evolved from the PICO1 (Patient, Intervention, Comparison, Outcome) arch interface, a method of arching
MEDLINE/PubMed that encourages the creation of a well-formulated arch.2 Starting from a clinical situation, a clinician is guided through the arch process by thinking
along PICO elements.
PICO arch was developed with the busy clinician in mind, interested in practising evidence-bad medicine, but unfamiliar with controlled vocabularies that could make
the arch more efficient. In an attempt to automate the entry of arch terms into PICO elements from a clinical question, we discovered that the ur could simply enter a
clinical question, then let the arch engine retrieve relevant journal articles. The step to allow the ur to inspect the correctness of PICO elements was omitted, however, a link
is provided to the PICO interface instead, so the ur can manually enter arch terms if
the arch results are deemed unsatisfactory,
Although PICO helps in finding recent evidence from MEDLINE/PubMed, some urs might still find arching for answers to clinical questions challenging. For some, it
might be becau it is time consuming to incorporate into their busy practice, difficult to learn, or perhaps, they may feel that it is not clinical enough and may not answer their question. They may find it more convenient to go directly to other subscription-only英语培训一般多少钱
3
resources. Although easier to arch and filtered for quality, the resources are not without disadvantages. They cover fewer journals than MEDLINE (Medical Literature Analysis and Retrieval System Online), and there is an even greater time lag between the publication of an article and its appearance in the databas than there is with MEDLINE. Other clinicians may just access resources that evidence-bad practitioners might consider less evidence-bad, (i.e. lesr validity) such as consulting a colleague,
or reading a textbook, or perhaps, they may forego arching altogether.
ask MEDLINE is intended for the clinician, rearcher, or the general public who want to simply ask a question and to skip the challenge of learning how to format it in manner
that will make the arching MEDLINE/PubMed efficient . It is a tool that allows the
ur to arch MEDLINE/PubMed using free-text, natural language query, just like one would in a clinical tting, or in a conversation. A ur enters a clinical question on a
Web browr, and then lets the tool retrieve relevant articles in MEDLINE/PubMed.
Links are provided to journal abstracts, full-text articles and related items. Moreover,
ask MEDLINE is formatted for easy viewing on a wireless handheld device so it can be ud while mobile, but will work equally well on a desktop computer.
We report our experience in developing ask MEDLINE, and an evaluation study on its potential to retrieve references, using published, evidence-bad resources.
Implementation
ask MEDLINE us a multi-round arch strategy. In the first round, the parr ignores punctuation marks and deletes words found on a “stop-word” list. The stop-word list includes PubMed stop words, and other words that we found by experience, to be detrimental to the arch. The parr, a PHP script, then nds the modified query to PubMed Entrez’ E-Utilities. The Extensible Markup Language (XML) file returned by E-Utilities indicates the category of each term in the query. Terms m
arked as “All Fields” denote that they are neither Medical Subject Headings (MeSH) terms nor MeSH Subheadings. The terms are checked to determine if they are found in a “MeSH
Backup vocabulary.” The backup vocabulary includes words other than MeSH terms,
such as MeSH descriptors, that are classified as “other eligible entries”. If an “All Fields” word is in the backup vocabulary, it remains in the query; if it is not, it is deleted. The remaining terms are nt back to PubMed, again through E-Utilities. Human and English language limits are always applied. If the journal retrieval count after the first round is between 1 and 50,000, the first 20 results are displayed in the ur’s browr and the arch process terminates. Further arches are dependent on the ur.
The arch may proceed to Round 2 under two conditions: 1) If no journals are found in
the first round, a result that could signify that the arch was too narrow (i.e., too many terms are arched, too many filters), the “All Fields” words are deleted from the query, even though they are found in the backup vocabulary. Only MeSH Terms and Subheadings remain (Round 2A.) 2) If the first round retrieval count is larger than
50,000 articles (an indication that the arch was too broad) the “All Fields” words removed during the first round (words not found in the backup vocabulary) are put back
4
geinto the query (Round 2B.) Round 2B arches contain all the MeSH terms (or MeSH
去日本留学的费用Subheadings) and “All Fields” words in the original question. The updated query from
512是什么意思either 2A or 2B is once again nt to Entrez E-Utilities. Retrieved journal articles are nt
to the ur.
Similarly, if the count returned from cond round is in the range of 1 to 50000, the
arch process terminates. If the cond round count is still equal to 0 (denoting that the
arch is still too narrow) another list of “No-Go Terms”, terms that when removed could
result in a successful arch is checked. Common MeSH abbreviations, acronyms and
words like, “method,” “affect,” and “lead” are examples of terms on the list. New terms
difficult用法
are continuously added to this list as they are encountered. The third round modified
query is once again nt to E-Utilities and the retrieved journal articles are nt to the
ur. A result of 1 to 50000 citations terminates the process and displays the first 20
得道多助失道寡助翻译articles.
If ask MEDLINE retrieves only one to four journal articles, a arch is automatically done
for related articles of the top two articles. All the articles (one to four previous) and the
first 25 related articles of the first two are retrieved. As in any of the previous steps, the
first 20 are displayed in the browr. In all the arch retrieval pages, a link is provided
kistafor the ur to manually intervene and modify the arch process through the PICO
interface. Links to related articles, full-text articles and abstracts are shown.
Since November 2002, the British Medical Journal (BMJ) has published a POEM
(Patient-Oriented Evidence that Matters) in every issue.3 POEMs are provided to BMJ by
InfoRetriever () ask MEDLINE was evaluated by comparing its accuracy to retrieve an article cited as a reference in a POEM (“gold standard”.) 3 Every
POEM has a question with a cited reference that is relevant to the question. We entered
every POEM question into ask MEDLINE, and for comparison, in Entrez, the integrated,
text-bad arch and retrieval tool for PubMed. New critically appraid topics (CATs)
from the University of Michigan, Department of Pediatrics Evidence-Bad Pediatrics
Web site were also ud.4 Unlike BMJ POEMs, some questions in CATs had more than
one cited reference.
The initial arch result was examined to determine if the reference cited in a POEM or
CAT was among tho retrieved. Subquent steps were taken if the reference article
cited was not: 1) If the initial arch retrieved journal citations, but not the specific
journals cited in a POEM or CAT, the titles and abstracts were scanned to find out if they
were relevant (deemed to answer the question.) If they were, related articles were
retrieved, and again evaluated to determine if they matched the cited reference. 2) If no
journal articles were retrieved, the question was rephrad, then arched again.
Retrievals were again examined for the cited articles and relevancy to the clinical
question. Overall efficiency was determined by the accuracy in retrieving a cited article
and relevance of citations retrieved for citations that did not match cited references.
5
道歉的英文
Results
Development of the arch tool
A simple, handheld-friendly arch interface was created where urs can enter free-text, natural lan
guage arches (Figure1.) The results are also provided in text-only mode, optimid for u by the mobile health care personnel using handheld devices. Links to abstracts, related citations in MEDLINE and full-text articles are provided. Full-text
links to journal publishers are not text-bad, and may require journal subscription or fees for viewing.
Evaluation of arch retrievals
Clinical questions in 95 POEMs and 28 CATs were arched. After first pass,
ask MEDLINE found 62% of the cited articles in POEMs, while Entrez retrieved clo to 14% (Table 1.) When related articles were arched, 11.6% more were found by
ask MEDLINE (8.4% in Entrez.) When three questions were rephrad, askMEDLINE,
but none in Entrez retrieved two of the specific cited references, although relevant references were found to one of the questions. For 20 questions, ask MEDLINE did not
find the specific cited reference, but it found journal citations that were deemed relevant and would b
e uful in answering the question. Entrez obtained citations for 16 (16.8%) questions that were considered relevant.
Overall, ask MEDLINE retrieved 72/95 exact matches of cited references (gold standard)
in POEMs, an accuracy of 75.8%, while Entrez’ accuracy was 22% (21/95.) If citations
that are not the same as tho cited in POEMs, but are relevant and considered satisfactory for answering the clinical question are included, ask MEDLINE’s total efficiency is 96.8%. Entrez’ total efficiency for finding specific and relevant citations for BMJ POEMs is 38.9% (21 specific and 16 relevant citations found.)
Although citations were retrieved for all POEM questions by ask MEDLINE, three arches did not find exact matches or relevant articles (3.1%), while Entrez’ results were not relevant (6.3%) for six questions. No citations were found for 52 (54% of total) questions by Entrez.
University of Michigan’s CATs yielded a similar total efficiency as POEMs, 96.3%,
while it was 14.3% for Entrez (Table2.) First pass retrieval was 64.2% (Entrez 3.6%) and citation retrievals for related citations was 10.7% (Entrez 3.6%.) Four of six questions rephrad added 14.3
% to the total efficiency of ask MEDLINE, while it added 7.1% to Entrez. Almost 7% of the arches retrieved relevant citations to rephrad or related articles, but none in Entrez. For CATs’ questions, ask MEDLINE found 89.2% of cited references, but 14.3% for Entrez. In 21/28 questions, Entrez did not provide a specific or relevant citation, but it was only for one question with ask MEDLINE.
Discussion
ask MEDLINE is part of a project to develop easy-to-u resources at the point of care
that has the functionality of an expert archer.5 Special consideration was given to
6
