Domain-specific journals and its social structure

From a scientometric perspective, science, metaphorically conceived as a knowledge space or knowledge landscapes, can be defined in terms of a network of scholars that produce a network of knowledge [35]. In the former case, the social function of science has long been recognized (e.g., by Thomas Kuhn): scholars produce and communicate scientific knowledge and this organized activity has the characteristics of a social process [36, 38]. More importantly, the patterns of interactions and communication within this social organization are tightly tied, rather than isolated, to the knowledge they produce [36].

An exemplary case is the role of journal editors as gatekeepers, with studies identifying editorial gatekeeping patterns [39, 40]. According with the Network Gatekeeping Theory, inspired by the work of Kurt Lewin, gatekeeping refers to the control in the flow of information [41, 42]. In the field of ED, this intellectual and social organization of knowledge can be seen in professional societies like the Academy of Eating Disorder, which since 1981 publishes the most renowned scientific journal: The International Journal of Eating Disorders. Within its editorial board, there are distinguished scholars that can act as gatekeepers to ensure quality control and that manuscripts published by the journal are in line with the aims and scope of it.

In sum, domain-specific journals have the goal of publishing information within the boundaries of their aims and scope, allowing the diffusion of specialized knowledge.

Domain-specific journals and its disciplinary organization

From a network perspective, specialty journals are also indicators of disciplinary organization [43], which exerts a non-trivial influence at both the global and local level of the network. To be more precise, if we visualize a network [e.g., 2, 44, 45], the local density of specialty journals evidence emerging patterns such as citation patterns by articles from the same journal or group of journals [43]. At the author level, these patterns reflect the local influence of specialty journals on scholars who adhere to their research tradition and their contributions help to advance a research agenda [46].

For example, domain-specific journals on ED often publish curated information from conferences [e.g., 47] or special issues about a specialized topic [e.g., 48], which commonly include a research agenda [48], setting the stage for future research. As we mentioned above, similar literature, such as special issues about ED published in other journals [e.g., 49], is not necessarily excluded in the analysis of domain-specific journals. Rather, such literature is commonly cited in documents from domain-specific journals and can be included in a citation analysis. Importantly, these citation patterns suggest that the former intellectual and social structure of knowledge constrains what is being studied in the future [46]. Thus, in the upcoming years, most of this specialized literature is expected to become an active research front [32], as evidenced by its high number of citations.

Finally, it is worth mentioning that the analysis of these patterns can reveal latent hierarchies and topological properties of journal networks. In fact, domain-specific journals can be identified through the study of the hierarchical organization of journal networks. When hierarchical network analysis is used to identify the capability of journals to spread scientific ideas, multidisciplinary journals are found at the top of the hierarchy, whereas more specialized journals are found at the bottom [50, 51]. Similarly, significant articles from a specific domain have unique topological properties that can affect the dynamic evolution of the network [52]. In sum, it is important to recognize the topological properties of networks and their latent hierarchies, both at the journal level and document level. In our case, focusing on domain-specific journals, it would be like zooming into the most central part (core) of the network topology to analyze its organization and distinctive features. Indeed, this approach is commonly employed, for example, when studying network subsets such as niches or communities in complex systems.

Domain-specific journals and complex adaptive systems

Domain-specific journals can also be comprehended from a complex systems standpoint, as the aggregation of the intellectual, social, and citation patterns outlined above. According to the Structural Variation Theory [53], the body of scientific knowledge can be conceived as a complex adaptive system (CAS). As such, it can be described and studied as a complex network with a series of characteristics like non-linearity, emergence, and self-organization; and a series of social, conceptual, and material elements that evolve over time [46]. Ideally, we must study CAS holistically to understand the properties of the system at the macrolevel [54]. In our case, this would require including all scholar literature on ED, which could be attempted using a query-based approach and employing ad hoc methodologies (e.g., iterative citation expansion) [45]. However, complex systems emerge from rules and behavior of lower-level components, and there is growing interest in understanding complexity from its simplest and fundamental elements and patterns [55, 56]. In our case, this can be accomplished by zooming into domain-specific patterns that emerge from the relational structure and organization of journals and papers [46], rather than focusing on the whole system which comprises all the scientific literature on ED.

This approach can be described in terms of modularity, a structural property of systems: the local density of specialty journals is indicative of a structural module or subsystem [57]. This property of complex systems is important because it recognizes, as we did above, the existence of subsets within networks. Indeed, scientometric studies usually attempt to detect communities based on the principle of modularity by grouping similar literature (i.e., clustering) [44, 58]. However, in the approach used in this study, rather than using bibliographic connections (e.g., through co-citation analysis) to detect domain-specific literature, we can use logical connections [59], to identify modules that operate as domain-specific representations [60]. In other words, domain-specific journals can be seen as clusters of articles that are logically linked because they all pertain to a given domain, which is explicitly stated in the aims and scope of the journals.

This modular organization has some advantages over others such as a hierarchy (e.g., Scimago categorization of journals) or a cluster obtained by literature partitioning algorithms. First, it has the advantage of reducing both complexity bias and hierarchical bias. The former is the tendency to assume and adopt a more complex system (the opposite to Occam’s Razor: prefer the simplest explanation), which means to analyze all ED literature. The latter assumes that behavior is directed in a hierarchical fashion, where a central authority passes instructions to all agents in the system [54]. Second, although it still recognizes a hierarchical structure composed by diverse classes of subsystems, it assumes heterarchy [43, 61], which means that both hierarchical and nonhierarchical elements can be present in a system; holarchy, which means that systems are composed of components that can be recognized as subsystems [62]; and glocal control, which means that local and global phenomena in a system are achieved by local actions [63]. In simple words, sampling a set of domain-specific journals reduces complexity without affecting assumptions such as a categorical hierarchy of journals.

The current study

To expand on previous studies [1, 2], the current study aims to answer the following research questions:

Data collection

The methodology workflow is presented in Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Workflow of the methodology used in this research study.

https://doi.org/10.1371/journal.pone.0278981.g001

First, in May 2020, a search of journals was performed in Scimago Journal Reports (SJR, https://www.scimagojr.com/), using the term “eating disorders”. In this step, the following five journals were identified: International Journal of Eating Disorders (ISSNs: 0276–3478, 1098-108X), European Eating Disorders Review (ISSNs: 1072–4133, 1099–0968), Eating Disorders (ISSNs: 1064–0266, 1532-530X), Eating and Weight Disorders (ISSNs: 1124–4909, 1590–1262), and Journal of Eating Disorders (ISSN: 2050-2974). The official website of each journal was then visited to confirm that the scope of the journal specifically includes the publication of research articles on eating disorders. It should be noted that the journal Advances in Eating Disorders (ISSNs: 2166–2630, 2166–2649) was not included because it was not found in SJR, it was published only between 2013 and 2016, it was incorporated into the journal Eating Disorders, and by the time of writing this article, it was not indexed neither in Scopus (https://www.scopus.com) nor in Web of Science (https://www.webofknowledge.com).

Next, also in May 2020, the Scopus database was chosen to retrieve the document records from the aforementioned journals. The election was made for no other reason than the capability of Scopus to retrieve several structured information (metadata, such as the abstract), and the file types for download are easy to manage, such as comma-separated values (CSV). Therefore, all document records published by these journals were searched in Scopus using the ISSN as the search term (e.g., ISSN (0276–3478) OR ISSN (02763478) OR ISSN (1098-108X) OR ISSN (1098108X)). A total of 8651 documents between 1981 and 2020 were retrieved (of which 7899 had an abstract): 4185 (48.38%) from the International Journal of Eating Disorders, 1540 (17.80%) from Eating and Weight Disorders, 1461 (16.88%) from the European Eating Disorders Review, 1072 (12.39%) from Eating Disorders, and 393 (4.54%) from the Journal of Eating Disorders. These 8651 documents included a total of 213,744 references. It should be noted that the International Journal of Eating Disorders is the oldest of these journals, established in 1981. The S7 and S8 Files provide the number of documents per year and per journal. The document records were downloaded from Scopus both as comma separated values (CSV) and as BibTex (http://www.bibtex.org/), and selecting all fields available (i.e., title, author, abstract, etc.). Due to copyright, the full text of all documents was not retrieved but rather their metadata (i.e, title, author, date, abstract), whilst the dataset shared online (https://doi.org/10.17605/OSF.IO/5YZVD) is the one obtained after the preprocessing procedures detailed below.

Analyses were performed using open software: R Statistical Software 4.0.3 (Bunny-Wunnies Freak Out) [65], and Python programming language version 3.9.1 (https://www.python.org/).

Bibliometric analysis and network analysis in R

The biblioshiny application from the R package bibliometrix [66] was used to preprocess the CSV file. Next, it was used to identify the most cited documents. Local citations (i.e., citations only from documents whithin the dataset), and global citations (i.e., citations made by any document from the whole Scopus database), were computed. Biblioshiny was also used for network analysis as described by Batagelj & Cerinšek [67], and Aria & Cuccurullo [66]. Regarding the network, it is defined as a pair of sets: a set of nodes or vertices and a set of edges (link between nodes) [68]. In this study, when authors were treated as nodes, a link would represent co-authorship or collaboration [see 69]. More precisely, the Louvain algorithm for community detection [70] was used to identify communities within the collaboration network. This algorithm identifies densely connected nodes within the network (i.e., communities) [e.g., 71]. It works unconstrained to automatically extract a number of clusters although it requires basic network parameters as input. These network parameters were: up to 100 nodes, a minimum of two edges by node, and the removal of isolated nodes. For network layout visualization, the Fruchterman & Reingold [72] algorithm was chosen. Finally, common centrality measures were calculated: betweenness, closeness, and PageRank. Betweenness centrality refers to “the frequency that a node is located in the shortest path between other nodes” [73, p. 772]. Closeness centrality refers to nodes that can easily reach others in the network, whilst PageRank, originally created to rank websites [74], has been used to rank authors because it takes into account the weight of influential nodes [75].

Topic modeling: Dimensionality reduction and matrix factorization

As can be seen in the workflow (Fig 1), once network analysis was finished, a series of steps (detailed in S5 File) were necessary to preprocess the dataset prior to topic modeling. Topic modeling refers to applying machine learning techniques to find topics by extracting semantic information from unstructured text in a corpus [76]. As we explain in S5 File, to this point we end up with a high-dimensional and sparse document-term matrix. In other words, we have many features (columns) each corresponding to a term in our corpus, and for a given document (rows) we have many columns with zero values meaning the term of that column is not in the given document. To deal with sparsity, we can perform dimensionality reduction to obtain a representation that effectively captures the variability in the data. In summary, dimensionality reduction can be categorized in feature extraction and feature selection; the former combines the original feature space into a new one, whereas the latter selects a subset of features [77].

As explained in S5 File, the term frequency (TF) and the term frequency-inverse document frequency (TF-IDF) were used as feature extraction for vectorization. Then, the following machine learning algorithms were applied for topic modeling: Latent Dirichlet Allocation (LDA) [78], Latent Semantic Analysis (LSA or Latent Semantic Indexing) [79], Hierarchical Dirichlet Process (HDP) [80], and Non-negative Matrix Factorization (NMF) [81]. LDA is a generative probabilistic model that decomposes the document-term matrix into a topic-term matrix and a document-topic matrix, and it is commonly used for topic discovering from a corpus [e.g., 82]. LSA utilizes a truncated Singular Value Decomposition for decomposition and can work efficiently on TF or TF-IDF sparse matrices. In a fully unsupervised framework, the HDP model is characterized by inferring the number of topics on its own. Finally, NMF is an alternative approach that implements the Nonnegative Double Singular Value Decomposition, an algorithm suitable for sparse factorization [83].

First, the GENSIM library [84] was used for topic modeling because it provides a way to calculate topic coherence, an index to compare models based on measures of segmentation, probability estimation, confirmation measure, and aggregation [see 85]. Therefore, based on a TF matrix, HDP, LSA, NMF, and LDA were performed in GENSIM and compared in topic coherence. Once identified the topic modeling algorithms with the highest topic coherence, scikit-learn [86] was used because it provides an Exhaustive Grid Search option for ensemble learning the models (i.e., automatically fine-tuning the parameters to find the most optimal). Finally, once the topics were extracted, a simple time series analysis was performed to visualize the changes over time in the topics found. This analysis consisted of simply plotting the number of documents for each topic across years, from 1981 to 2020.

Bibliometric analysis

The top cited documents were all from the International Journal of Eating Disorders. As noted above, this journal is the oldest one (it started in 1981), and it has the largest number of articles per year, with the exception of the year 2019 when it was outperformed by the Eating and Weight Disorders journal (see S7 and S8 Files). The majority of top cited documents were related to the development of instruments for the assessment of eating disorders or the course and outcome of eating disorders. For example, we can see in the results the most common instruments used for the screening of eating disorders, as well as the evaluation of its core symptoms: Eating Disorder Inventory (EDI), Body Shape Questionnaire (BSQ), Dutch Eating Behavior Questionnaire (DEBQ), and Eating Disorder Examination Questionnaire (EDE-Q). These instruments are widely used to screen the general population, as well as in clinical settings, together with more recent instruments [88]. It should be noted, however, that in clinical practice settings the use of instruments for the diagnosis and the different phases of the treatment process is not necessarily widespread [89, 90]. To reduce this gap, some authors suggest to provide assessment training and/or assessment guidelines for mental health professionals and general practitioners in primary health care [91, 92]. This can help obtain a comprehensive clinical assessment, particularly of individuals with higher risk such as young adolescents with restrictive Anorexia Nervosa [93]. The instruments mentioned above are reliable measures, and they could be used online for a quick screening or session by session for ongoing monitoring, although further research is necessary [e.g., 94–96].

The rest of most cited documents include important review articles on epidemiology (Hoek & van Hoeken, 2003, in Table 1); the course and outcome of eating disorders (Berkman, Lohr & Bulik, 2007; Strober, Freeman & Morrell, 1997; in Table 1); and the diagnosis of Binge Eating Disorder (Spitzer et al., 1992, 1993, in Table 1). These results are similar to previous studies in which measurement methods (including instrument development), epidemiology, and review articles were the most common type of document [8, 9].

Finally, the large number of articles on the diagnosis of Binge Eating Disorder, which was not fully recognized as a mental disorder in the Diagnostic and Statistical Manual of Mental Disorders (DSM) until its fifth edition [4], reveal that the recognition of Binge Eating Disorder as an own disorder took several years. To reach expert consensus in a shorter time, eating disorder professionals should pay special attention to emerging eating problems, such as Orthorexia Nervosa [97].

Network analysis

The network analysis identified eight clusters with 96 authors. Previous studies have examined the network of authors in the field in terms of network statistics such as number of edges or network density [2]. By contrast, this study provides a more fine-grained network analysis, identifying experts and group of experts in the field of eating disorders. As seen in the results section, the majority are distinguished authors with contributions dating back to the early 1980s.

The author with the largest betweenness centrality was Ross D Crosby (Sanford Center for Biobehavioral Research, United States), followed by James E Mitchell (University of North Dakota, United States) which has the largest value in PageRank. Authors with high betweenness centrality can act as both enablers and gatekeepers of information flow between communities [75]. Moreover, it has been found that authors with high betweenness centrality establish more collaborations than those high in closeness centrality [75]. In summary, the results of centrality measures can help to identify experts in the field of eating disorders, particularly authors that can quickly reach other authors in the network (high in closeness), act as gatekeepers (high in betweenness), or relate to influential others (high in PageRank).

Regarding the clusters identified by the network analysis, in the same cluster of Ross D Crosby and James E Mitchell are found other renowned authors like Daniel Le Grange (University of California, San Francisco, United States), Stephen A Wonderlich (Sanford Center for Biobehavioral Research, United States), and Carol B Peterson (University of Minnesota, United States). Among the most relevant results of collaboration of this cluster we can find studies on the ecological momentary assessment of eating disorders [98], the psychometric properties of the EDE-Q [99], and the diagnosis of Binge Eating Disorder [100].

The second largest cluster includes authors like Cynthia M Bulik (University of North Carolina at Chapel Hill, United States), Walter H Kaye (University of California, San Diego, United States), and Katherine A Halmi (Weill Cornell Medical College, United States). The results of their collaboration include studies related to the phenotypic characterization of eating disorders, such as the International Price Foundation Genetic Study, a multisite study that included a large sample of patients with eating disorders and their families [e.g., 101].

Finally, the third largest cluster includes authors like Janet Treasure (King’s College London, England), Ulrike Schmidt (King’s College London, England), and Tracey D Wade (Flinders University, Australia), which are widely recognized by the Maudsley Model for Treatment of Adults with Anorexia Nervosa (MANTRA) [102, 103]. Interestingly, this is the only cluster that includes collaborations with authors from non-English speaking countries, more specifically from Spain. Examples of these collaborations include studies resulting from the Wellcome Trust Case Control Consortium 3 (WTCCC3) and the Genetic Consortium for AN (GCAN) [104], and other studies with clinical samples in Spain [e.g., 105].

On the other hand, the results reveal the importance of multisite studies that strengthen collaboration and originate in relevant outcomes for the prevention and treatment of eating disorders. Research groups could look for opportunities to collaborate in multisite studies and strengthen both their interdisciplinary and transdisciplinary collaboration, and their collaboration with less common partners such as stakeholders and policy makers [106, 107]. By establishing these integrative and strategic collaborations we can promote translational research, and thus helping to reach broader public health goals [108].

Topic modeling

The combination of TF-IDF and NMF provided meaningful results, identifying 10 topics. After labeling these topics based on the first 10 keywords and their respective weights, we can see that most of the research on eating disorders done in the past 40 years has focused on their prevention and treatment. Interestingly, the time trend analysis of these topics revealed a noticeable change in the first lustrum of the 1990s. Whereas during the early 1980s the study of clinical groups (topic 5) was the most dominant topic, from the mid-1990s, this topic was surpassed by the study of risk factors of eating disorders (topic 1). This indicates an increasing interest for the prevention rather than solely the treatment of eating disorders. This result is consistent with the historical shift that occurred in the United States when in 1992 the Institute of Medicine (IOM) Committee on Prevention of Mental Disorders was created [109]. Then two years later, a report on reducing risk factors for mental disorders and promoting a preventive approach in research was published [110]. As expected, this shift had echo in several scholars at the time, became a research front, and relevant publications started to include more information on the prevention of eating disorders, including a special issue [111], book chapters [112], and progressively entire books [113]. It is important to note that this historical shift, as well as later others like in 2017 [114], were favorable, because in other cases like obesity, it took more time to focus on its prevention due to different issues, including the pressure of the weight loss industry and its commercial interest [115].

Another interesting finding was that the outcome of the treatment of eating disorders (topic 6), is the second most important topic of 2013, and this finding has important aspects to discuss. First, the surge of state-of-the-art machine learning algorithms provide several opportunities to build intelligent systems for precision medicine. Thus, the treatment course and outcome of eating disorders can be more personalized, guided, and enhanced with the help of predictive technologies and intelligent systems [e.g., 116]. Second, as suggested elsewhere [117], the advantages of technology can be particularly relevant for certain age groups like adolescents, and when a digital intervention is employed [118]. In summary, treatment outcome is currently an important topic, and future studies can deploy digital interventions and machine learning algorithms for a more precise treatment planning.

Limitations and conclusions

Although this study has strengths, such as using data and code that allows the reproducibility of the results, readers should consider some limitations. First, the analysis of most cited documents is for all the time span, and more recent highly cited documents are underrepresented. Moreover, the journal Advances in Eating Disorders was not included due to indexing issues. Nevertheless, this study provides the code and a detailed procedure to allow researcher to perform further analyses, such as document co-citation analysis. Future studies can also evaluate the Mexican Journal of Eating Disorders (Revista Mexicana de Trastornos Alimentarios, ISSN 2007-1523), which has published articles primarily in Spanish [119]. Second, the network analysis included close to 100 scholars mostly with a long trajectory in the field, and this can be a limitation in representing more younger scientists or newcomers [2]. Future studies can focus on a larger number of scholars and apply different techniques in network analysis, such as other community detection techniques [e.g., 120]. Finally, the results of topic modeling suggested a solution of 10 topics out of up to 30 topics solution models tested. Although there is not a universally accepted approach to establish the number of topics, this study relied on several strategies, including ensemble learning, to automatically fine-tune the parameters of the machine learning algorithms, stability, and heuristic approaches [121]. Future studies can try other machine learning algorithms and techniques to retrieve topics [121].

In conclusion, this study analyzed 40 years of research on eating disorders, identified the most cited articles, networks of collaboration, experts in the field, and the 10 major topics in the field.