Multi-Objective Optimization for Clustering of Medical Publications

Multi-Objective Optimization for Clustering of Medical Publications Asif Ekbal Sriparna Saha Indian Institute of Technology Patna, Bihar, India asif@iitp.ac.in sriparna@iitp.ac.in Diego Moll´a Department of Computing Macquarie University, Sydney NSW 2109, Australia diego.molla-aliod@mq.edu.au K Ravikumar Indian Institute of Technology Patna, Bihar, India ravi.mc12@iitp.ac.in Abstract Clustering the results of a search can help a multi-document summarizer present a summary for evidence based medicine (EBM). In this work, we introduce a clus- tering technique that is based on multi- objective (MOO) optimization. MOO is a technique that shows promise in the areas of machine learning and natural language processing. In our approach we show how MOO based semi-supervised cluster- ing technique can be effectively used for EBM. 1 Introduction Evidence Based Medicine (EBM) urges the med- ical doctor to incorporate the latest clinical evi- dence available at point of care (Sackett et al., 1996). However, the amount of published clinical evidence is enormous. PubMed,1 for example, in- dexes over 23 million citations, and the amount is growing every day. There are systematic reviews such as Cochrane’s reviews that distill and summa- rize the information relevant to a particular topic, but often the doctor needs to access the primary literature, especially for cases that are rather in- frequent and do not have systematic reviews ded- icated to them, when dealing with particular seg- ments of the population, or when the patient has si- multaneous conditions (“comorbidity”). A search to PubMed can easily return hundreds of results, and finding specific information from that sea of information is time-consuming. To help the doctor’s need to find the evidence, it has been proposed to cluster the search results ac- cording to the different topics present in the clini- cal answer (Shash and Moll´a, 2013). The motiva- tion for this is that answers to a clinical question usually have several distinct parts, each of which 1http://www.ncbi.nlm.nih.gov/pubmed Which treatments work best for hemorrhoids? 1. Excision is the most effective treatment for thrombosed external hemorrhoids. [11289288] [12972967] [15486746] 2. For prolapsed internal hemorrhoids, the best definitive treatment is traditional hemorrhoidec- tomy. [17054255] [17380367] 3. Of nonoperative techniques, rubber band ligation produces the lowest rate of recurrence. [1442682] [16252313] [16235372] Figure 1: PubMed IDs of documents relevant to the answer to a clinical question. is backed by a distinct set of published evidence. For example, as shown in Figure 1, the documents that answer the clinical inquiry which treatments work best for hemorrhoids? published in the Jour- nal of Family Practice2 can be grouped into three clusters, one for each suggested treatment (exci- sion, hemorrhoidectomy, rubber band ligation). We therefore propose to cluster all the docu- ments relevant to a clinical query into clusters. Given a collection of clinical questions, the doc- uments of each question represent a separate clus- tering task. In this paper, we present a method that uses multi-objective optimization techniques to cluster the results. Section 2 gives a brief survey of clustering in general and within EBM. Section 3 introduces the general framework for the multi-objective opti- …

Content truncated to 500 words