Performance evaluation of unified medical language system^®'s synonyms expansion to query PubMed

The most important tool to access the medical literature is the PubMed search engine, which allows access to more than 20 millions of biomedical citations. The major part of these citations comes from the MEDLINE bibliographic database, which uses the MeSH thesaurus for indexing [1]. Other citations, i.e. OLDMEDLINE, out of scope or recent citations, are not indexed at the time of user query [2]. The most comprehensive way to find citations in MEDLINE is to use the MeSH thesaurus.

Because one third of Medline queries are performed by members of the general public [3] and furthermore because most health professionals [4] are not aware of this thesaurus, they run free-text queries, as they do when using Google™. This allows searching the entire PubMed collection but does not at all exploit the indexing work produced by National Library of Medicine (NLM) building the MeSH thesaurus and indexing millions of citations. Consequently, the US National Center of Biotechnology Information has developed several techniques (Automatic Term Mapping (ATM)) to map end-user queries to the MeSH thesaurus and other search field descriptors (e.g. author's name, publication's name, etc.) [5]. The first ATM aim is to improve information retrieval in structured information: searching indexes (mostly MeSH terms used to index citations in MEDLINE) instead of only the free text. Almost nothing is done to enhance the search for recent citations (not indexed). This is a limiting factor because 1) these citations contain the most recent scientific discoveries and 2) they are the first returned by PubMed, which displays recent articles first, by default.

Although PubMed ATM query is continuously improved, a recent review [6] has counted 28 different entities that have devoted themselves to develop Web tools for helping users to quickly and effectively search and retrieve relevant publications on MEDLINE. This highlights the need for alternative ways of searching the medical literature. Thirion et al. [7] have shown that it is possible to improve ATM's performance, mainly in precision (tools available with the Doc'CISMeF search engine http://​www.​cismef.​org) using MeSH synonyms. The aim of this paper was to propose an extension to this previous optimization, using Unified Medical Language System^® (UMLS) synonyms, and to assess its performance.

Table 2 summarizes results for the 43 queries we had to perform in order to obtain 20 citations for 20 queries. For the other 23 queries, q5 query did not produce any results: enhancing query using UMLS synonyms did not add any further results. The new strategy led to a heterogeneous 23.7% mean increase in non-indexed citation retrieved (from 0 to 9,876 new citations retrieved). None of the three tested factors (number of sons in MeSH hierarchy, of MeSH synonyms and of UMLS synonyms) were significantly correlated with the number of citations retrieved or the precision.

For the 20 studied MeSH Descriptors, inter-rater agreement was poor: multi-rater's kappa was 0.34. Results of relevance evaluation are summarized in Table 3. The mean precision for fully relevant citation was 48.4% CI_95% = [45.8-50.9] but this number does not reflect discrepancies between evaluators: three evaluators (BT, GK and NG) found full relevance around 40% (43.7%, 36.7% and 37.4%, respectively) and one (LR) found 75%. Results are somewhat better for partially relevant citations but have a similar pattern, LR's evaluations were often more relevant than other evaluations: mean partial precision was 59.8% CI_95% = [57.3-62.2]. BT, GK and NG found a precision of about 50% (50.1%, 51.7% and 48.2% respectively) whereas LR found 88.1%. Differences between evaluators were significant (p < 0.001, Friedman test). There was also a significant difference of precision depending on the MeSH term (data not shown, p < 0.001, Chi² test): for 8 MeSH term the full relevance precision was higher than 0.5, for 8 MeSH terms the partial and full relevance precision was less than 0.5.

Enhancing information retrieval is one possible use of UMLS [12]. The new strategy led to a slight increase in non-indexed citation retrieval (23.7%) for a precision very similar to those observed in previous reports studying PubMed performances: Thirion et al. [7] showed a precision of 54.5%; Lu et al. [13], for a normal use of PubMed, found a mean rank precision for the 20 top results between 40% and 55%.

Nevertheless, this study has some limitations: First, the absence of a control group to make a comparison led to difficult interpretation of results. However, consistency with literature review suggests that there was no major bias. Second, we used queries based on one MeSH term from the "disease" tree (C). However, would the results be similar for other MeSH tree terms, queries including several MeSH terms or queries including MeSH terms and keywords? Third, there is great variation in the results of the expansion proposed here between queries. The three factors tested were not significantly correlated with precision, but a qualitative assessment of the results was manually performed:

(a) Some UMLS synonyms provide very good results (e.g. "hepatoma" for "liver neoplasms", "Nephropathy" for "Kidney Diseases"), probably because they are very similar to a son of MeSH Descriptor.

(b) Some UMLS synonyms are ambiguous acronyms that generate a lot of noise (e.g. TB for tuberculosis).

(c) Some MeSH descriptors correspond to frequent confounding factors (e.g. "hypertension", "obesity"). Results of retrieved citations are adjusted based on these factors but they are not the real subjects of the citations (mean precision for fully relevant citations: 21.1%, 23% respectively).

We have also tried to explain the number of newly retrieved citations using the q5 query, which varies from 0 to 9,876 (for Diabetes Mellitus; see Table 2).

(a) The number of UMLS synonyms greatly varies from 0 to 38, with a median of 10 (data not shown). This natural source of variation was not confirmed by correlation tests. The difference must be more qualitative:

(b) Some UMLS synonyms do not provide any added value in information retrieval (e.g. all the synonyms finishing by ", NOS" will not provide any citation).

Fourth, the relevance assessment was performed on title and abstract alone, but not with the full text of the article. Although this could have introduced a bias in this study, it seems to us more pragmatic as most end-users select the relevant citation based on title and abstract alone. Lastly, the poor inter-rater agreement measured here (kappa = 0.34) suggests that we do not really know what we are measuring, even if it is common for this type of study [14]. This poor kappa score, and the surprising distribution of results, only highlights differences between users. The improvement proposed here is probably not of interest for some users but may be of interest for others. Based on this study, we have implemented the following three procedures to query MEDLINE via PubMed in the following tool InfoRoute, French Infobutton (URL: inforoute.churouen.fr) [15]:

Different types of users should use these three procedures. Users expecting the most exhaustive results, even at the cost of some noise, should use the latest one. This type of users wants to maximize the recall.

Lu et al. [6] reviewed 28 different ways to access MEDLINE citations. The search strategy we propose could possibly be the 29^th. However, when compared to other teams' strategy to improve PubMed information retrieval, the ones developed by our team modify the ATM and then are applicable in the PubMed interface. In fact, there is no need to integrate and update the MEDLINE bibliographic database in our information system.

Considering the huge number of citations retrieved by each q3 query (frequently more than dozens of thousands, data not shown), the increased number of recent citations retrieved may not lead to an important increase in recall. Nevertheless, the proposed strategy is based on the following assertion: a citation that is not indexed with a MeSH term does not have to be retrieved whatever the semantic expansion was used.

Based on this, the new strategy will only retrieve new citations not belonging to MEDLINE that represent more than ¾ of PubMed citation. We observed a 23.7% increase in recall for the citations aimed by the new strategy, which is not insignificant for the users, especially if they are searching for recent scientific advancements. This improvement mainly concerns new citations (82% of the citations retrieved by q5 have been published less than 4 months earlier). Furthermore, these citations, ranked first by PubMed, may be of great interest for PubMed users who frequently do not read more than the top 20 answers [16].

In contrast to PubMed, we assumed that when end users search for a disease name in PubMed, they do not add synonyms because of laxity or unawareness. It could be useful to add the son's preferred terms, son's entry terms and son's UMLS synonyms to the query with "ORs". This would eventually lead to an increase in recall and in proportion of queries retrieving additional citation (20 on 43 for this study) and a decrease in precision. However, it would drastically increase query size and resources needs, which are already quite substantial.

The expansion of queries using UMLS' synonyms may not be of interest for all PubMed users, but could be quite useful when seeking for exhaustivity (review, meta-analysis, etc.) as well as when searching for the latest scientific citations. This study highlights the need for specific search tools for each type of user and use-cases.