Clinical Information Extraction Using Word Representations

Clinical Information Extraction Using Word Representations Shervin Malmasi ♣ Hamed Hassanzadeh ♦ Mark Dras ♣ ♣ Centre for Language Technology, Macquarie University, Sydney, NSW, Australia ♦ School of ITEE, The University of Queensland, Brisbane, QLD, Australia shervin.malmasi@mq.edu.au, h.hassanzadeh@uq.edu.au mark.dras@mq.edu.au Abstract A central task in clinical information ex- traction is the classification of sentences to identify key information in publications, such as intervention and outcomes. Sur- face tokens and part-of-speech tags have been the most commonly used feature types for this task. In this paper we eval- uate the use of word representations, in- duced from approximately 100m tokens of unlabelled in-domain data, as a form of semi-supervised learning for this task. We take an approach based on unsuper- vised word clusters, using the Brown clus- tering algorithm, with results showing that this method outperforms the standard fea- tures. We inspect the induced word rep- resentations and the resulting discrimina- tive model features to gain further insights about this approach. 1 Introduction Evidence-based Medicine (EBM) is an approach to enhance clinical decision making by leverag- ing currently available evidence. The rationale be- hind EBM is that clinicians can make more judi- cious decisions with access to abundant clinical evidence about a particular medical case. This ev- idence is sourced from research outcomes which can be found in medical publications accessible via online repositories such as PubMed.1 Al- though millions of publications are available, find- ing the most relevant ones is cumbersome using current search technology. Additionally, the rapid growth of research output makes manual analysis and synthesis of search results unfeasible. This has given rise to the need for methods to auto- matically extract relevant information from publi- cations to support automatic summarization (Has- 1http://www.ncbi.nlm.nih.gov/pubmed sanzadeh et al., 2015). This is an emerging re- search area that has begun to attract increasing at- tention (Summerscales et al., 2011). This information extraction is generally per- formed at the sentence level on the paper ab- stracts (Verbeke et al., 2012). Scholarly publica- tions usually follow a common rhetorical structure that first defines the problem and research aims by introducing background information. They then describe the methodology and finally the out- comes of the research are presented. Abstracts, as the summary of the reported research, gener- ally have the same structure. This information, which can be considered as scientific artefacts, can usually be found in the form of whole sentences within the abstracts. More specifically, the arte- facts in the clinical research domain have been categorized as Intervention, Population or Prob- lem, Comparison, and Outcome. This is known as the PICO scheme (Richardson et al., 1995). An- other proposed approach to formalise the rhetori- cal structure of medical abstracts is the PIBOSO model (Kim et al., 2011), a refined version of the PICO criteria. It contains six classes, rather than four: (i) POPULATION: the group of individuals participating in a study; (ii) INTERVENTION: the act of interfering with a condition to modify it or with a process to change its course; (iii) BACK- GROUND: …

Content truncated to 500 words