1. Automatic selection of references for the creation of a biomedical literature review using citation mapping
Houcemeddine Turki
Faculty of Medicine of Sfax, University of Sfax, Sfax, Tunisia
INTRODUCTION
 I only consider the 100 search results that were cited the most in 2016 when building the co-citation network and the direct citation network of the topic of the literature review.
This will be automatically done using VOSviewer (Fig. 3 and 4) [11].
Literature reviews are important for scientists and particularly biomedical ones as these scientific publications can be based on a limited number of verifiable references to provide a complete overview of every topic a biomedical scientist would like to learn about [1, 2]. However, the selection of the significant references to be used in a biomedical literature review can take time when manually done as the number of articles about a given topic that are picked using a Medline search query can be very high.  Scientists tried to make algorithms and workflows to automate this step [1, 3]. These methods are based on applying the NLP techniques on the titles, the abstracts and the keywords of the references [4, 5] or on citation mapping and analysis techniques [6, 7, 8]. As a continuation of these efforts to automate the selection of references for the creation of a biomedical literature review, I define in this research work a novel workflow that integrate direct citation network analysis and co-citation network analysis to select the references for a biomedical literature review from a list of MeSH-based Medline query results.
Fig. 4: Co-citation Network
Fig. 3: Direct Citation Network
SELECTION OF THE SECONDARY REFERENCES
PROPOSED WORKFLOW
As shown in [12], the automated method with high precision that can be used to retrieve secondary references that explain the background or the findings of the primary considered references for a given literature review is snowballing technique. Snowballing technique consists of retrieving important papers about the topic having a direct or indirect citation link with the primary references [12] (This is evident as these papers are the ones that mostly explain the historical evolution, the adjustments and the scientific basis of the findings involved in the primary references [7, 12]).  I build the direct citation network of the initially considered papers about the topic using CitNet Explorer (Fig. 5) [13]. Then, I simply consider the papers that are included in the independent subnetworks of the direct citation network involving the primary references and that have a significant weight within the direct citation network as the secondary references.
The proposed workflow is the following:
Retrieval of the initial considered set of references using Web of Science and Medline
Choice of the primary references through the analysis of the citation network and the co-citation network
Choice of the secondary references through snowballing
RETRIEVAL OF THE INITIAL SET OF REFERENCES
Fig. 1: Proposed Workflow
RETRIEVAL OF THE INITIAL SET OF REFERENCES
The research question will be converted to a MeSH-based query before proceeding to using a specialized software or directly by searching MeSH keyword database available in An explanatory example for this step is provided in Fig. 2
After the MeSH-based query is formulated, I use Web of Science to retrieve the search results of the query from Medline.
Use of drugs to cure osteoporosis in Europe
becomes
"Osteoporosis/drug therapy"[Mesh] AND "Europe"[Mesh]
Fig. 2: Sample conversion of the title of a literature review into a Medline MeSH-based query
SELECTION OF THE PRIMARY REFERENCES
As explained in [9], when two articles are cited together in many papers, it is likely that these papers are similar. Consequently, it is useless to cite both articles in a literature review.  The papers with the best weight in the co-citation network of the topic = the main papers that report most of the findings about topic [6].
Similarly, papers that are used as references in many papers about the topic are probably the main ones that were used to build all the findings about the topic [7, 10].  The papers with the best weight in the direct citation network of the topic = the main papers that explain most of the findings and the historical development of the topic [6].
Primary references are the papers having the best weights in the co-citation network and the direct citation network of the topic of the literature review [6, 7, 10].
As insignificantly cited search results are cannot have a weight in the
. co-citation network and the direct citation network of the topic of the literature review.
Fig. 5: Historiographic Direct Citation Network
FURTHER INFORMATION
Further details about the scientific basis, the precision, the speed and the running matters of the workflow can be found in the Proceedings Paper of the work available online.
CONCLUSION
In this paper, I propose a citation analysis-based complete workflow to drop irrelevant or useless papers from the initial set of considered references of a biomedical literature review (Generally search results of a Medline MeSH-based query) using a combination citation analysis and mapping techniques.
Future directions of this research can be the substitution of the paid citation database used in this work (Web of Science Core Collection) by a free and wider citation database (Google Scholar), the amelioration of the speed of the workflow, or the creation of the software that automatically applies this workflow.
MAIN REFERENCES
[1]
G. Tsafnat, P. Glasziou, M. K. Choong, A. Dunn, F. Galgani and E. Coiera, Systematic reviews, vol. 3, no. 1, p. 1, 2014.
[2]
L. Bornmann and H. D. Daniel, Journal of Documentation, vol. 64, no. 1, pp , 2008.
[3]
G. Tsafnat, A. Dunn, P. Glasziou and E. Coiera, BMJ, vol. 346, p. f139, 2013.
[4]
D. Zhang, J. Lei and A. K. Robinson, "A hybrid approach for automating citation screening process," in Abstracts of the 21st Cochrane Colloquium, Québec City, 2013.
[5]
C. J. Groot, T. Leeuwen, B. W. J. Mol and L. Waltman, Paediatric and perinatal epidemiology, vol. 29, no. 6, pp , 2015.
[6]
K. W. Boyack and R. Klavans, Journal of the American Society for Information Science and Technology, vol. 61, no. 12, pp , 2010.
[7]
E. Garfield, Journal of Informetrics, vol. 3, no. 3, pp , 2009.
[8]
M. W. Bruner, K. Erickson, B. Wilson and J. Côté, Psychology of Sport and Exercise, vol. 11, no. 2, pp , 2010.
[9]
H. Small, Journal of the Association for Information Science and Technology, vol. 24, no. 4, pp , 1973.
[10]
H. Keathley-Herring, E. Van Aken, F. Gonzalez-Aleu, F. Deschamps, G. Letens and P. C. Orlandini, Scientometrics, vol. 109, no. 2, pp , 2016.
[11]
N. J. Van Eck and L. Waltman, Scientometrics, vol. 84, no. 2, pp , 2010.
 [12]
M. K. Choong, F. Galgani, A. G. Dunn and G. Tsafnat, Journal of medical Internet research, vol. 16, no. 10, p. e223, 2014.
[13]
N. J. van Eck and L. Waltman, Journal of Informetrics, vol. 8, no. 4, pp , 2014.