Aggregating semantic information nuggets for answering clinical queries

Abstract

In this paper, we address the issue of answering PICO 1 clinical queries formulated within the Evidence Based Medicine framework. Answering clinical questions gives raise to numerous challenges among wich term ambiguity and relevane estimation based on the distribution of the query facets in the documents. The contributions of this work include (1) a new algorithm for query refinement based on the semantic mapping of each facet of the query to a reference terminology and (2) a new document ranking model based on a prioritized aggregation operator that leverages the importance of each facet with regard to a candidate relevant document. The effectiveness of our PICO-based search approach is empirically evaluated using a clinical retrieval collection including 423 queries and more than 1.2 million of medical abstracts from PubMed. The experimental results show that our approach for PICO query answering significantly overpasses state-of-the-art document ranking models.

Semantic Information Retrieval on Medical Texts

“…(1) Local context: first, an initial retrieval run is performed to build a list of ranked documents in response to the user's query. Most of the proposed QE methods are based on traditional document ranking models such as the language model [38,115,153,181,198] and the probabilistic model [151,201]. Then, terms contained in the top-ranked documents are extracted using a blind or pseudo-relevance feedback approach.…”

“…(2) Global context: candidate expansion terms are identified based on their explicit vs. implicit one-to-one or one-to-many semantic relationships with the query terms or the terms issued from the local context (if [38] Local & Global Specialised Multiple (MEsH, SNOMED, GO, ICD) Zhu and Carterette [198] Local & Global Specialised Multiple (MeSH, medical collections) Limsopatham et al [89] Local & Global Specialised Multiple (MeSH, MedDRA 40 , DOID 41 ) Oh and Jung [115] Local & Global Specialised (Medical collections) Multiple & General (Medical collections) Shondi et al [153] Local Specialised (MeSH) Single (corpus and top ranked documents) Martinez et al [100] Global Specialised (UMLS) Single (knowledge resource) Wang and Akella [170] Local & Global Specialised (UMLS) Single Znaidi et al [201] Local & Global Specialised (UMLS) Single Soldaini et al (a) [152] Local & Global General (Wikipedia) Single Soldaini et al (b) [151] Local General (Wikipedia) Single Xu et al [181] Local & Global Specialised (MeSH) Single Balaneshinkordan…”

“…Term Ranking/Selection Function Shondi et al [153] Enhanced term query frequency if term 𝑡 is associated to relevant types (e.g., disease) Martinez et al [100] PageRank term score over the query graph Relationships between terms are provided by the resource Wang and Akella [170] Maximum likelihood estimation of term assignment to relevant types Znaidi et al [201] Term score based on a recursive propagation algorithm through the query graph Soldaini et al (a) [152] Likelihood of being health-related Soldaini et al (b) [151] Likelihood of being health-related Xu et al [181] Learning-to-rank -based score Table 5. Main term ranking and selection functions used in AQE methods using a single resource.…”

“…To identify candidate expansion concepts, most previous work in the area of medical IR used either a specialized ontological resource (e.g., UMLS [170,181] or MeSH [153,201]) or a raw specialised textual corpus (e.g., health-related pages of Wikipedia [152]). Basic ontological-based approaches commonly identify candidate concepts: related concepts having generalization/specialization or part of relationships with the query concepts; terms generated from the local context [152,201].…”

“…To identify candidate expansion concepts, most previous work in the area of medical IR used either a specialized ontological resource (e.g., UMLS [170,181] or MeSH [153,201]) or a raw specialised textual corpus (e.g., health-related pages of Wikipedia [152]). Basic ontological-based approaches commonly identify candidate concepts: related concepts having generalization/specialization or part of relationships with the query concepts; terms generated from the local context [152,201]. For example, with respect to the UMLS resource, the query 'cancer' is likely to be expanded with the synonymous concept 'malign tumour' and the query 'osteoporosis' could be expanded with the (preferred) term of concept 'Boniva' since UMLS states that 'Bonivia' is the treatment of 'osteoporosis.…”