Due to the increasing popularity of new research in medicine this study was conducted to determine recent research trends of predictive, preventive and personalized medicine (PPM). We identified the terms relevant to PPM using own search engine based on neural network processing in PubMed database. We extracted initially about 15000 articles. Then we carried out the statistical analysis for identifying research trends. The article presents the results of solving the problem of evaluating research topics at the level of thematic clusters in a separate subject area. An approach based on the analysis of article titles has been implemented. Identification of terms, connections between them and thematic clustering were carried out using the free software VOSViewer, which allows to extract terms in the form of noun phrases, as well as to cluster them.
2[0000-0003-3557-009X], 3[0000-0002-6357-7929], Bibliometric mapping helps transform most of the metadata of publications into maps or visualizations, from which useful information can be obtained through post-processing.
Bibliometry is one of statistical methods to analyze the mass of literature and to
reveal historical development [
For example, Zhu & Guan (2013) [
VOSviewer uses the VOS display technique (visualization of similarities) and is
freely available (www.vosviewer.com). Maps are generated from a sharing matrix. The
similarity matrix is calculated using the measure of the strength of the association [
When creating maps based on a text corpus, the user can choose between binary and full counts. When choosing a binary count, only the presence or absence of the term in the document is considered. In the case of a complete count, all occurrences in the document are considered.
Content analysis is one of the areas of bibliometric analysis and it includes the
identification of trends based on words (Huffman et al., 2013 [
Gelman and Unwin [
The general process of visualization of scientific maps and their subsequent analysis consisted of the following stages. As a principle for constructing the matrix, a network based on jointly occurring keywords was chosen
1. Obtaining a semi-structured amount of information from information sources containing document annotations, as well as information about authors, their affiliation, publication date, keywords, citation information (universal information retrieval system Dimensions; specialized PubMed database).
2. Pre-processing of data to improve the quality of the generated map (the formation of groups depending on the period).
3. Construction of scientific maps and their visualization.
4. Analysis of scientific maps, identification of the most intensively developing topics, analysis of temporal, statistical aspects.
In this study, the PubMed database created by the National Center for Biotechnology Information (NCBI) in the United States was identified as the first base for collecting
A Scientometric Analysis of Publications Related to Predictive Medicine 3
information. The search for standard bibliometric processing was performed based on
the terms "predictive" and "personalize (s) e" "preventive" "medicine" with the logical
operator "AND". We measured the number of publications in the field of PPM
depending on the affiliation of the first author. The output of publications at the country level
was estimated using PubMed tools [
To isolate key terms using the Word2Wec method, about 15,000 articles (titles and
annotations) were selected from the PubMed database, containing the terms
"predictive" and "personalize (s) e" in their titles. To extract keywords from the headlines, we
used baseline Medline / PubMed database for all years. The annual baseline is released
in December of each year [
Algorithm for finding trends when processing a corpus of natural language texts: 1. Initially, an expert creates a dictionary of key terms, consisting of keywords and phrases in a normalized form. This is a dictionary De or main dictionary. 2. Next, a temporary dictionary-template of new words and phrases is created Dn. 3. Then an array of statistics is created to analyze the neighborhood of keywords.
S = { Dei{ Dnk, Rk}} (1) Dei is an element of key terms vocabulary,
Dnk – new word (not a stop word or a word from the main dictionary De), Rk – a frequency of occurrence of a new word Dnk in the vicinity of the key term Dei. At first, its value is zero.
4. If a key term is found, then an analysis of its neighborhood is performed. The neighborhood of the word is determined by the size of the sliding window. The words of the analyzed text are normalized. If a word (not a stop word) is found in the vicinity of a keyword it is not included in the main dictionary, it is first checked whether the array S contains a pair of values { Dei, { Dnk }}. If there is no such pair of values, then it is added to the dictionary. Then the frequency value (Rk) for a given pair of values, increases by one. Each new element Dnk can become either a key term or part of it.
5. After processing the text corpus and filling in the S array, the most rated candidates for key terms are selected. Minimum rating threshold (Rmin) the new term is determined by the expert. нового термина определяет эксперт. All elements of the array S with a rating less than Rmin are not considered.
6. Next, candidates for a new key terms are generated. If there are several new words for the same keyword with the same Rk, then there is a possibility that the new term could be more than one word. The system generates a possible new multi word term. At the same time, candidates for key terms are generated with already approved elements (from the main dictionary De) and possible key terms found in the vicinity of the keyword, but already without it.
7. As a result of the analysis of the values of the array S and approval by the expert new key terms are added into the dictionary. New key terms with the maximum rank determine development trends in this field of medicine.
The “Dimensions” scientific information database was used as the second database. The search was carried out by keywords in the annotations "predictive preventive personalized medicine" for 2008-2020.
VOSviewer version 1.6.15 was used to construct and visualize scientific maps based on the data obtained. The program allows you to carry out scientific mapping based on scientometric analysis (frequency of co-occurrence of keywords).
The construction of a graph of interconnected pairs of objects is based on multidimension scaling (MDS), which is a means for visualizing pairwise connected graph vertices displayed on a plane. The method is used when two or more dimensions need to be examined for data analysis. MDS is used to construct bibliometric maps, which can include either co-citation of sources or co-occurrence of terms (objects, authors, documents, journals, keywords). The number of joint occurrences of elements within a given neighbourhood i and j is denoted cij. ci is a frequency of element i. If we have a set of n vertices, we want to build a graph of connected elements (for example, vertices that indicate the joint appearance of elements, or joint citation of documents), then sij will mean the strength of the associative connection between vertices i and j (Van Eck & Waltman, 2009) and represents.
= ∑ ≠ = 2
= 1 ∑
2 They call this measure as proximity index. Otherwise, we can say that cij means the total number of connections of the vertex i and m means the total number of connections in the network.
When constructing a graph, two types of similarity can be used: direct and indirect. When using the indirect type of similarity to build a graph, direct comparison of elements is used, when using the indirect type of similarity, vectors of elements are compared. MDS allows to reduce the dimension of the graph. The method works in a similar way as t-distributed Stochastic Neighbor Embedding (t-SNE).
There are several options for implementing MDS, including Metric
multidimensional scaling (mMDS) [
We analyzed trends in the field of predictive and personalized medicine (PPM) for the period from 1940 to 2020.
To search in PubMed the keywords “predictive medicine”, “personalized medicine”, “preventive medicine” were used. The collocation “preventive medicine” at first was (2) (3) (4)
A Scientometric Analysis of Publications Related to Predictive Medicine 5 appeared in 1857. The collocation “predictive medicine” at first was appeared in 1918. The collocation “personalized medicine” at first was appeared in 1952.
From the graph (see Fig. 1) we can see that predictive medicine, personalized
medicine and preventive medicine experienced ups and downs in popularity. There is the
increase of publications in these areas in last decade, apparently due to the activity of
the European Association of Predictive, Preventive and Personalized Medicine
(EPMA), which was established in 2008 [
s 1200 n ito 1000 a c i lb 800 u fpo 600 r eb 400 m iN 200 0 0 3 6 9 2 5 8 1 4 7 0 3 6 9 2 5 8 1 4 7 0 3 6 9 95 95 95 95 96 96 96 97 97 97 98 98 98 98 99 99 99 00 00 00 01 01 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 Y2ea20r 20 preventive medicine predictive medicine personalized medicine Fig. 1. The number of PubMed publications in the fields of preventive, predictive and personalized medicine from 1950 to 2020. 3.2
VOSviewer carries out the extraction of keywords in a matrix, the structure of which forms in the form of a network by linking terms according to the calculated link strength. Bond strength was defined as the total number of occurrences of a term in pairs with other terms.
The first step involved mapping terms extracted from an array of articles indexed in Dimensions for the period 2008–2020. For further visualization, a list of 85 keywords was formed, which included terms from article titles. In VOSviewer, the option "create map from text corpus" was selected. The counting method was set to binary counting. This means that each concept is counted only once per annotation, regardless of how many times it actually appears in a given annotation. VOSviewer's algorithm automatically excludes 40% of least significant terms. In total, 51 terms were used for mapping out of 4979 terms (see Fig. 2).
Stage 1 involved rendering a text corpus of 15,000 titles identified during the preparation phase. These names were extracted from the PubMed database (see section 2).
The purpose of this step is to create an overview of concepts within a given body of literature.
Stage 2 consisted of two parts. Part 1 was aimed at extracting key terms using special software tools. 6 Aida Khakimova et al.
The method is based on the sequential linear division of the sentence into various phrases. The statistics of various options for parsing phrases in the corpus were calculated using a floating "window" - a numeric hyperparameter that shows how many words to the left and right of the central one is its environment. In this way, more than 9200 derivatives and phrases were extracted.
Part 2 focused on mapping key terms with overlapping windows (see Fig. 3). Part 3 focused on answering a research question about how software tools can help identify interesting ideas.
An additional visualization was performed in which the threshold for the minimum number of occurrences was set to 5 and the number of concepts included in the visualization was set to 123 (see Fig. 4). This step was done with a binary count. Non-specific terms were excluded (Asia, October, USA, New York, 21st century, etc.). The list of terms / concepts defined in VOSviewer can be used as an initial contextual mapping template. Contextual analysis can also help to identify potential relationships between two or more concepts.
For example, let us select the first cluster. It includes the terms assessment, clinical, discovery, evaluation, innovative approach, personalized medicine approach, predictive, preventive, rare disease, translation, translational.
We found these terms in our list of 9200 terms and phrases. The term «predictive» is central to the cluster in terms of the number of links. We built a context map of terms for this set of terms, considering the results of VOSviewer and our statistical processing of the corpus of articles. 4
As discussed in the previous section, 123 concepts were included in the visualization in Phase 1 of the analysis. The VOSviewer map gave 14 conceptual clusters. After exploring the concepts in each cluster, we selected 1 cluster to build the contextual conceptual mapping.
Cluster 1 focuses on topics related to innovative approaches in predictive preventive personalized medicine (shown in table 1). Cluster 2 is devoted to topics related to per- sonalized management and application of biomarkers. Cluster 3 can be referred to as Preventive Medicine and Cardiology Perspectives. Cluster 4 focuses on personalized diagnosis and medicine in oncology. Cluster 5 brings together topics related to the paradigm and strategy of personalized medicine. Cluster 6 deals with Omics in the treatment of rheumatoid arthritis. Cluster 7 focuses on topics related to personalized preventive medicine in cancer treatment. Cluster 8 includes issues of genetics and genomics in the PPM. Clusters 9 to 12 cover topics related to PMP in the management of chronic non-cancer diseases (multiple sclerosis, diabetes, Parkinson's disease COPD). Clusters 13 and 14 focus on innovative PPM in oncology, including Non-small-cell lung carcinoma (NSCLC). Data from 8 cluster have not included in the table 1.
The above list of research areas may already be of interest from a scientific landscape perspective. At this point, you can go back to the beginning of the process and develop a search strategy based on one cluster.
The starting point for contextual conceptual mapping (see Fig. 5) was a list of concepts of the first cluster.
The list was supplemented with additional concepts that proved to be important. The question of whether a term or concept was considered potentially significant was determined in terms of the frequency of its occurrence in titles and annotations (as determined by frequency analysis).
Figure 5 shows the results of the extended contextual analysis. This map can serve as a starting point for a deeper review of the literature. 5
We have presented a method for analyzing publications and identifying trends in predictive medicine. The proposed method allows to search for new terms in preventive medicine based on the neighborhood approach.
Experts took part in identifying new trends. The use of software to automatically highlight trends significantly reduces the time it takes to generate new terms. To process the corpus of selected articles, one expert would need several months, while using the proposed software and considering the work of experts, this work was completed in 2 weeks.
The modified word2vec method was used by the authors to highlight key terms and to build forecasts. The diagrams are built using the MDS method. The search and analysis of information was carried out in several electronic libraries using their analytical mechanisms and visualization tools. A limitation of the current study is that the analysis relied on information provided by the authors in the titles. Future research may include full text analysis. Two contextual conceptual term maps were created in the prototype. These maps were created to provide an overview of research topics and to identify promising directions.
This work is supported by RFBR, grants 18-07-01111, 18-07-00909, 18-07-00225, 1907-00455 and 20-04-60185.
References:
A Scientometric Analysis of Publications Related to Predictive Medicine 11