Evolution of primary care databases in UK: a scientometric analysis of research output

Objective To identify publication and citation trends, most productive institutions and countries, top journals, most cited articles and authorship networks from articles that used and analysed data from primary care databases (CPRD, THIN, QResearch) of pseudonymised electronic health records (EHRs) in UK. Methods Descriptive statistics and scientometric tools were used to analyse a SCOPUS data set of 1891 articles. Open access software was used to extract networks from the data set (Table2Net), visualise and analyse coauthorship networks of scholars and countries (Gephi) and density maps (VOSviewer) of research topics co-occurrence and journal cocitation. Results Research output increased overall at a yearly rate of 18.65%. While medicine is the main field of research, studies in more specialised areas include biochemistry and pharmacology. Researchers from UK, USA and Spanish institutions have published the most papers. Most of the journals that publish this type of research and most cited papers come from UK and USA. Authorship varied between 3 and 6 authors. Keyword analyses show that smoking, diabetes, cardiovascular diseases and mental illnesses, as well as medication that can treat such medical conditions, such as non-steroid anti-inflammatory agents, insulin and antidepressants constitute the main topics of research. Coauthorship network analyses show that lead scientists, directors or founders of these databases are, to various degrees, at the centre of clusters in this scientific community. Conclusions There is a considerable increase of publications in primary care research from EHRs. The UK has been well placed at the centre of an expanding global scientific community, facilitating international collaborations and bringing together international expertise in medicine, biochemical and pharmaceutical research.


INTRODUCTION
Big data (analytics) refer to the aggregation and interrogation of-high volume, high velocity, high variety-data sets so as to reveal new, non-obvious, information and patterns. 1 This field is advancing because of technological and scientific developments in information infrastructure and digitisation. 2 For governments, opening up the data sets states hold about their citizens is believed to have, through computational and algorithmic analyses, a disruptive and transformative effect on knowledge. 3 In the UK, big (open) data have been at the forefront of research activity and policymaking. Termed as one of the eight great technologies, 4 UK has embraced the big (open) data movement more than many other developed countries (eg, USA, Australia, France). 3 One area of particular relevance to big data analytics is healthcare.
In UK, the National Health Service (NHS) is organised around primary care and, unless there is an accident or emergency, whenever citizens would like to use the NHS they have to go through their primary care physician, known in the UK as a general practitioner (GP). From there, they can be referred to a specialist at a hospital if necessary. Secondary care clinicians can then feedback information to GPs. Since the vast majority of the population (98%) is registered with a general practice, GPs act not only as the main gatekeepers for the NHS but also as important custodians of a longitudinal electronic health record (EHR). 5 There are now many ongoing primary care databases of Strengths and limitations of this study ▪ First study to perform a scientometric analysis of research output from primary care databases of electronic patient records. ▪ We analysed articles published from 1995 to 2015 in order to explore the historical breadth and growth of this type of research. ▪ The analysis is limited to articles and structured data retrieved from the Scopus database. ▪ Some latest articles and related citations might not have appeared in Scopus when the data set was extracted.
anonymised patient records in UK that can be used for healthcare research. These population-based databases contain data originating from routine general practice. Some newly established databases and research platforms of linked EHRs include ResearchOne 6 and CALIBER. 7 While there are more than 9600 general practices in UK that could potentially contribute data to these databases, 8 it is usually 6-10% of these practices that do so. Such databases are usually used for cross-sectional surveys, case-control or cohort studies and for epidemiological, drug safety, clinical and healthcare usage research purposes. They rely heavily on individual general practices voluntarily contributing data via the propriety clinical systems they use to maintain these patient records. The records are usually anonymised or pseudoanonymised at source by allocating a unique number to each patient to allow for the updates of the records as well as for their linkage to other data sets, such as national mortality, national cancer registration and hospital records as well as with socioeconomic, ethnicity and environmental data sets. Access to these data sets is usually granted after scientific and ethics review and can be tailored to customer requirements. In this study, we examined the research output of three such databases that are well established in the research community and have contributed to a substantial number of scientific studies and publications. These are the Clinical Practice Research Datalink (CPRD), 9 The Health Improvement Network (THIN) 10 and QResearch. 11 The CPRD (formerly known as the General Practice Research Database) is a not-for-profit research service funded by the NHS National Institute for Health Research (NIHR) and the Medicines and Healthcare products Regulatory Agency (MHRA). It is owned by the UK Department of Health and contains the records of 11 million patients (4.4 million active) from 674 general practices. 5 There is a service cost associated with the preparation of the requested data. Unlike the other services described below, CPRD does not extract data from a particular propriety clinical system. Any general practice can contribute data after a data sharing agreement with the software supplier. Importantly, it is the only database accessible online. 12 The THIN database contains the health records of 12 million patients from around 600 general practices that use the Vision clinical system by In Practice Systems (INPS). IMS Health can provide access to data, for example, via a yearly sublicense to an academic institution. THIN is the only database that can provide access to data for for-profit companies. QResearch is a research service located at the University of Nottingham. Its database contains the health records of 18 million patients from 1000 general practices that use the EMIS clinical system. Only academics employed by a UK university can have (in site) access only to a sample of the data set (maximum 100 000 patients) that is sufficient to answer a specific research question or hypothesis. As this research service is not-for-profit and entirely self-funded, there is usually a fee to be paid to cover the cost of the data extraction.
The strengths of these databases lie in their size, breadth, representativeness of the UK population, longterm follow-up and data quality. 5 They include good information on morbidity and lifestyle, prescribing, preventive care, current standards of care and interpractice variation. 13 Since they are continually (and automatically) updated, they are ideal for researchers to discover and monitor healthcare trends as well as the effectiveness of new interventions and treatments, with minimum cost. They are increasingly linked to secondary care and mortality data sets. In contrast, their weaknesses include the fact that data are extracted from propriety clinical systems developed for patient management and not for healthcare research. There are issues of missing data (eg, from healthy patients), variable definitions for diagnoses and incomplete secondary care data (eg, hospital admissions). Wider health data (eg, treatment adherence, over the counter medication) and data about subpopulations (eg, prisoners, homeless people, refugees, travellers) are not captured adequately. 5 Information governance and informed consent procedures around the data sharing of EHRs for research are still considered complex. 14 These databases also require considerable clinical and scientific expertise, as well as technical capacity in data management to support research. When selecting a particular data resource for an observational study, researchers have to consider several other factors, such as the population covered and its geographical distribution, data capture and latency, linkage with other resources, privacy and security, quality and validation. 15 Nonetheless, these databases are highly regarded within the research community since they have proved their value in helping researchers reach definitive answers in various healthcare debates of considerable public interest, particularly where other types of research have produced contradictory evidence. For example, in 2004 researchers from UK and Canada proved beyond doubt that measles-mumps-rubella (MMR) vaccination is not associated with autism in children. 16 In contrast to expensive, time-consuming, and unrepresentative (of the population) traditional randomised trials, 17 large-scale and randomised observational (comparative) evaluations of treatments and medications are minimally obtrusive for clinicians and patients and can support faster turnaround times for pragmatic evidence useful in clinical practice. 18 The aim of this study was to perform a scientometric analysis of articles, published from 1995 to 2015, which have used data from at least one of these primary care databases. This empirical, semiautomated, method of quantitatively analysing a large number of publications provides a reliable and objective examination of the current status and trends as well as the structure and dynamics of this scientific field. 19 20 In this way, policymakers, research funding bodies but most importantly new researchers entering this field can have a general overview of its knowledge base and an indication of what kind of network features, research activities and topics of interest are driving it. 20 21 To the best of our knowledge, this is the first study of a systematic mapping of primary care databases research output.

METHODS
In this study, Elsevier's Scopus database (http://www. scopus.com) was selected as the source of structured data on articles. This database covers more scientific articles than other databases (eg, Thomson Reuters Web of Science) and has the advantage of providing advance export functionality of structured data, including full citation information, abstracts, keywords and references. The results were then limited to articles, articles in press, conference papers, reviews, book chapters and short surveys. Notes, letters, editorials and errata were excluded from the analysis. From there, we compared the resulted records with the bibliographic lists maintained by these databases 22-24 so as to include articles that could not be retrieved using the above search queries. Data cleansing included the removal of duplicate records and records that were missing essential information for the analysis (eg, article title, journal). The fields of authors, year, source title, affiliations, author keywords and document type were used for the analysis.
The final bibliography retrieved from Scopus was imported to Table 2 Net 25 to extract networks of authors and contributing countries. It was then imported to Gephi 26 where the ForceAtlas 2 algorithm 27 was used to visualise the structural proximities for the communities of authors and contributing countries. The VOSviewer (V.1.6.3) 28 software was used to visualise bibliometric networks and densities 29 of frequent terms and journals. All other statistical analyses were performed using Microsoft Excel. We used the Journal Citation Reports ( JCR) Science Edition 2014 to extract impact factor values for the identified journal titles.

RESULTS
A total of 1891 papers from 1995 to 2015 were included in this bibliometric and scientometric analysis. The results are presented below.

Publication and citation trends
The literature related to the 3 primary care databases in England increased gradually from 7 papers in 1995 to 171 in 2015 (table 1). We estimated their compound annual growth rate (CAGR), for the years 1995-2014, to be 18.65%. The vast majority of papers were published in English across 425 different sources (16.76% CAGR for 1995-2014). In total, these papers have already been cited 73 929 times. There is, however, a small percentage of 1.16% (n=163) papers that have not yet been cited yet. The average citation per year is ∼3.52.
We explored the distribution of publications by document type. This is presented in table 2 to identify the preferences of scholars using these databases in their research to share knowledge. The vast majority of scholars prefer to publish the findings of their research through journals, particularly as original articles (96.5%).
Next, we analysed the distribution of papers based on the academic discipline in which they have been categorised by Scopus (table 3) and by which each paper may be attributed to more than one subject area. 30 Since we analysed bibliographic data based on published research using primary care databases, it comes as no surprise that the vast majority of papers are under the medicine category. There is, however, a considerable number of papers (∼25%) under the categories biochemistry, genetics and molecular biology and pharmacology, toxicology and pharmaceutics, which indicates an emphasis on the use of these databases for the study of medications. It also indicates the potential interest in these databases from the pharmaceutical sector.

Most productive institutions and countries
For a deeper insight into contribution patterns and scientific impact, we first identified the top 10 institutions (by number of papers) authors have used as affiliation and then we analysed citation patterns (table 4). We also analysed authors' affiliations based on the country of their institution. For this, each publication was assigned to its authors' respective affiliated countries so as to identify the network of multinational collaborations. The distribution of the top 10 contributing countries is presented in table 5. Finally, we visualised the network of contributing countries using Gephi. We ended up with a network of 29 nodes and 175 edges (figure 1). Each node represents a country, while its size denotes the country's degree and the colour the number of papers.
The thickness of interconnected lines (edges) denotes the number of coauthored papers between the countries. The majority of the most productive institutions are universities. Top universities include the University of Nottingham, Boston University, University College London (UCL), the London School of Hygiene & Tropical Medicine and the University of Utrecht. Apart from these academic institutions, a research unit in Spain (CEIFE) and the MHRA in the UK are involved in primary care databases-based research. In this table, we also report the medians along with the IQRs. From this, it seems that scholars from CEIFE, University of Pennsylvania and Boston University are coauthors in publications that are highly cited compared to the other institutions in this list. Switzerland, Canada, Sweden, Germany, France and Italy had no institution among the top 10 list, although they were ranked among the top 10 productive countries. Most papers are published by scholars from UK (63.56%), followed by the USA and Spain. With the exception of USA and Canada, most of the productive countries are in Europe. What is particularly interesting in these two tables is that scholars in institutions from the USA and Spain produce not only a great  number of publications but also publications that are widely recognised by this scientific community in terms of citations.
Taking into account the measurements of weighted degree, clustering, eigenvector centrality and betweenness centrality (table 6), we observe that once again the UK, followed by USA, is placed at the centre of this scientific community. With the highest degrees of all measurements, institutions from this country are the most well-connected and authoritative ones, facilitating the linking between institutions in other countries. In table 7 What is also of particular importance in terms of scientific impact is that the 10 most cited papers identified (see table 8) have not been published in journals in this  list. However, by performing a (full counting) analysis of cocitation links in VOSviewer (figure 2) for journals cited in the Scopus data set (minimum number of cita-tions=10) we see that most of this list is represented here (blue-lowest density, red-highest density).

Most cited papers
Next, we focused on the top 10 papers 31-40 and calculated the total count of citations for each paper (table 8)  Of the 10 papers, 8 are single country papers, while none were singled authored. Again, the USA has a considerable presence in this list, producing papers that are highly cited. In addition, many of the highly productive authors identified (table 10) were also found in this list.

Authorship patterns and networks
Authorship distributions varied from single to a maximum of 155 authors-for a study about the feasibility of international collaboration to evaluate, based on a common protocol, the risk of Guillain-Barré syndrome following pH1N1 vaccination. 41 In total, there were 9385 authors involved in the 1981 papers during 1995-2015. In table 9, we can see that more than three-quarters of all papers were published by three or more authors. Only 1% of papers were written by a single author, while five papers did not have any authorship details. Almost a quarter of all papers was published by four authors, which have been widely cited across this scientific community. This indicates the high degree of expert collaboration in this field, that is necessary in analysing millions of primary care records. Table 10 provides the ranking of the top 10 scholars, first, in terms of research productivity based on the overall number of coauthored papers. While, generally, most scholars are from the UK, the Director of the Spanish Centre for Pharmacoepidemiologic Research (CEIFE) 42 is the scholar with the most published research from these primary care databases. Also, there are researchers in this field who do not necessarily come from the academic environment. The pharmaceutical sector is actively involved in knowledge production from electronic primary care records.
Considering only those scholars who have coauthored at least two papers in this data set, the analysis suggested a network (figure 3) with 1261 nodes and 6186 edges. Here, each node represents an author, while its size denotes the number of author's papers. The interconnected lines (edges) denote the coauthored papers between those authors. For better visualisation, we limited the number of minimum degrees to 5 (maximum degrees=145). After a modularity measurement, to identify community structure, 43 we observe some established collaborative teams (clusters with different colours) around specific and highly productive scholars in the analysis of data from primary care databases also found in table 10. We also observe a new (blue) cluster around the lead statistician for THIN 44 -one of the three primary care databases studied.
Taking into account the measurements of weighted degree, clustering, eigenvector centrality and betweenness centrality (table 10), results indicate a cluster placed at the centre of this scientific community. With the lowest degree of clustering and the highest degrees of all the other measurements, its prominent scholar is the most well-connected, facilitating, more than any other scholar, linking between other scientific clusters and scholars.  7

Open Access
What is also particularly interesting is the fact that some of the scholars (and their institutions) in this list are affiliated, to a certain extent, to these databases, having served or currently acting as their founders, directors, lead scientists or members of their scientific committee. 45 46 For example, lead scientists from the Boston Collaborative Drug Surveillance Program in Boston University were among the first who developed the technical and scientific capacity of these databases in pharmacoepidemiological research. 47 48 Research topics We conducted a keyword analysis to identify important topics of published research. For this, we first extracted from the bibliographic data set 5813 unique keywords as indexed by Scopus 30 to base our analysis on more complete indexing information compared to authors' keywords. We retrieved the top 30 keywords for two specific categories: medical conditions and medications/substances (tables 11 and 12). Next, we created a (full counting of ) term co-occurrence density map in VOSviewer (figure 4) by building a text corpus out of the title and abstract fields in the bibliographic data set (minimum number of a term occurrence=10). In this way, we were able to identify topics that not only appear more frequently in the literature, but that were also strongly related to each other, forming clusters of topics. Blue indicates a low density of terms and red indicates the highest density of terms. In many cases, the density map represents the frequency of indexed keywords in tables 11 and 12. Clearly, smoking, diabetes, cardiovascular diseases, mental illnesses, psoriasis, obesity, pregnancy and cancer as well as medication and substances that can treat these medical conditions, such as aspirin, insulin, antidepressants and non-steroid anti-inflammatory agents (NSAIDs), have been of great interest for scholars using EHRs in primary care.

DISCUSSION
This study identified the leading institutions, countries, authors, journals and topics as well as their networks of published research that have used primary care databases in the UK to extract and analyse data from EHRs. There is a growing production of such papers which indicates the interest of a global and highly collaborative scientific community in this field and also the knowledge and insights that can be gained for healthcare improvement. Publication output increased from 7 papers in 1995 to 171 by October 2015 (18.65% CAGR for 1995-2014). It may be worth noting that by performing a similar, limited to the UK, search in Scopus for the same period and with the keyword 'primary care' we found a 10.83% CAGR, which shows the increase in research conducted from these databases outstrips the field more generally. The vast majority of publications (96.5%) were journal articles. While this research field can be located, generally, in medicine, biochemical and pharmaceutical developments seem to be equally important, aimed at addressing widespread medical conditions, such as diabetes, cardiovascular diseases, mental illnesses, smoking, obesity and cancer.
The UK has been well placed in this scientific field. This is partly due to the fact that there is now more than 30 years of data available in GP information systems. 49 The investment in developing primary care databases from EHRs for research purposes has placed the country at the centre of a network of collaborations across the globe, bringing together international expertise for the analysis of ever-expanding and increasingly interlinked clinical data sets from primary, secondary and tertiary care. Access to these data sets has also allowed researchers and institutions from other countries to develop their own programmes of research to answer important clinical questions. Six of the most productive institutions are located in the UK, and 63.56% of publications were authored by scholars affiliated with this country, followed by the USA. Interestingly, the top institutions were not exclusively universities. Among them, we can find a research unit in Spain (CEIFE) and the executive agency in UK that funds and runs the CPRD database (MHRA), while one of the most productive scholars is affiliated with a pharmaceutical company. This signifies the great interest of various actors, from academic, governmental and private sectors, in research with primary care databases.
The geographical trend can also be observed from the location of journals with the most published papers. Six of the top 10 journals are published in the UK, followed by the USA. The journals with the most papers published included Pharmacoepidemiology and Drug Safety, BMJ, British Journal of General Practice and British Journal of Clinical Pharmacology, which signifies the great interest of scholars in using data from EHRs for pharmaceutical research. This is partially because one of the oldest sets of routine information collected by GP practices in the UK and made available by these databases is drug histories. 47 Regarding restrictions on access to research outputs, only four journals in this list are fully open access. This may limit access to knowledge to researchers and members of the public that cannot afford subscription costs. Interestingly, it is the more established journals in medicine, such as JAMA, Lancet and NEJM, that have published some of the most cited papers in this bibliometric data set and enjoy a high level of cocitation activity. Keyword analyses show that smoking, diabetes, cardiovascular diseases, mental illnesses, psoriasis, obesity, pregnancy and cancer constitute the main topics of research activity using EHRs in primary care. Often, this research concentrates on developing algorithms to identify risk of occurrence of a particular disease. Researchers are also interested in investigating medications that can treat these medical conditions, such as aspirin, other NSAIDs, insulin and antidepressants.
For the vast majority of publications, authorship varied between three and six authors, indicating widely collaborative, international, efforts to promote research in this field. Coauthorship network analyses showed that the lead scientists, directors and founders of these databases were found, to various degrees, at the centre of clusters in this scientific community, highlighting their invaluable contribution to knowledge production. As Azoulay et al 50 have demonstrated in their study about eminent researchers and the vitality of a field, the development of coauthorship networks and clusters of collaborators in newly established scientific domains might be useful to boost research productivity. On the basis of each database's data access requirements, their established researchers appear to have a fundamental role in facilitating and promoting international collaborations for institutions and countries. Importantly, they have a clear and in-depth understanding of the kind of research activities these databases can support in terms of data quality, structure and EHR coding practices. As these databases are expected to open up in the future to more stakeholders from various disciplines around health and as universities prepare to incorporate training in data science skills (eg, statistics, biomedical informatics, biology and medicine) 51 into their clinical curricula, so as to nurture the next generation of clinical investigators, 52 these established researchers could promote quality, reliable and ethically appropriate scientific research 53 from complicated and highly contextual data sets.
Our study has the typical limitations of a scientometric study. We analysed articles published in a period of 20 years in order to explore the historical breadth and growth of research from electronic primary care records. However, this analysis is limited on structured data retrieved from one bibliometric database of peerreviewed literature. Therefore, only articles published in journals in its index were analysed. Also, some of the latest articles and related citations might not have been retrieved at the time of the search, which might explain the decrease in the number of publications and citations particularly from 2010 onwards. It was beyond the scope of this quantitative study to assess the scientific quality and the socioeconomic impact of the large number of publications analysed here. These studies have deployed a range of study designs across many subfields of primary care research and with various research findings. Our main objective was restricted to assessing one aspect of academic impact and research quality, that is, patterns and trends in research outputs. 54  research could focus on the wider academic and socioeconomic impact of these studies by examining the relationship between publications, citation patterns and collaborations with the development of new scientific methods in the field or of new medical products and healthcare services.
In conclusion, output of primary care research from EHRs has consistently increased since their development. The development of these databases in the UK has placed the country and affiliated academic institutions at the centre of an expanding global scientific community, facilitating international collaborations and bringing together international expertise in medicine, biochemical and pharmaceutical research.
Contributors PV and ST conceived and designed the study. PV collected and analysed the data. PV and ST interpreted the findings, drafted the manuscript and approved the final manuscript for submission. Disclaimer The views expressed are those of the authors and not necessarily those of the EC, NHS, the NIHR or the Department of Health.
Competing interests None declared.
Provenance and peer review Not commissioned; externally peer reviewed.
Data sharing statement No additional data are available.
Open Access This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http:// creativecommons.org/licenses/by/4.0/ Figure 4 Term co-occurrence density map.