Exploring OpenAlex Topics in the Field of Health Professions

Exploring OpenAlex Topics in the Field of Health Professions Brendan Clay, Molly Frost, Naomi Mattli-Lewis, Leah Reavie School of Information Studies (ÉSIS) University of Ottawa ISI 5302: Knowledge Organization Fall 2025 Taught by Dr. Stefanie Haustein Teaching assistant: Chantal Ripp 1 Introduction OpenAlex is an open source document-based bibliographic database created to replace Microsoft Academic Graph (MAG), which was discontinued by Microsoft in 2021 (Priem et al., 2022). Developed by the Open Science nonprofit OurResearch, it offers an open alternative to proprietary databases, such as Elsevier’s Scopus, Clarivate’s Web of Science, and Digital Science’s Dimensions. Housing upwards of 240 million works, OpenAlex uses a hierarchical structure to classify its works into domains ( n = 4), fields ( n = 26), subfields ( n = 252), and topics ( n = 4516). This structure is an expansion upon MAG’s structure and was initially selected to facilitate the migration of works from MAG to OpenAlex; as OpenAlex’s coverage has expanded, its structure is being modified and refined to better support users (OpenAlex, 2024). OpenAlex topics, which offer the most granular level of classification, are a recent addition to the system. When a work is added to OpenAlex, it is assigned to a single topic, and is thus also classified under that topic’s subfield, field, and domain. The hierarchy is organized such that a given work will only belong to one category at each level of classification. For instance, if an article is assigned to the topic Healthcare Systems and Challenges , it will also be assigned to the subfield General Health Professions , the field Health Professions , and the domain Health Sciences . Given that topics are a new component of the OpenAlex hierarchy, their utility, as well as any issues with their coverage, have yet to be assessed. To this end, we are interested in evaluating the topics that fall under the field of Health Professions and its eleven subfields. This paper touches first on literature concerning OpenAlex to establish the environmental context of our research. It then covers our evaluation of the classification system, which includes a breakdown of topic distribution among subfields, an investigation into topic naming conventions, and our evaluation of indexing for articles assigned to Health Professions topics. Our aims are to better understand the current OpenAlex landscape, and to highlight areas that help or hinder its function as a tool for open research. 2 Literature review and environmental scan The precursors to the current OpenAlex hierarchy were OpenAlex concepts , which were derived from MAG’s fields of study (FoS). During the migration of works from MAG to OpenAlex, the number of FoS was greatly reduced by removing all FoS containing fewer than 500 works and automatically reclassifying them into the surviving FoS, which were relabelled as concepts. Approximately 715,000 FoS in MAG were reduced to around 65,000 concepts in OpenAlex (Scheidsteger and Haunschild, 2023), demonstrating a decrease in the number of distinct categories, but an increase in the breadth of each category. Although the number of concepts was a drastic decrease from the number of FoS, OpenAlex initially retained the poly-hierarchical structure of MAG. Concepts were distributed across six hierarchical levels (Priem et al., 2022), and concepts in lower (child) levels could be descended from multiple higher (parent or ancestor) levels. Zafar (2025) presents instances of concepts with multiple parents, and notes that just over 70% of concepts had two or more parent concepts at the 1 highest level of the hierarchy. Scheidsteger and Haunschild (2023) also note that descending down the concept hierarchy did not necessarily increase granularity; in fact, the second and third levels contained the most concepts, and the fifth (lowest) level only contained about a tenth of the total number of concepts. OpenAlex’s present structure is decidedly different from MAG’s organization, and from the concept hierarchy presented above. In 2024, OurResearch restructured OpenAlex’s classification system based on work by researchers from the University of Leiden's Centre for Science and Technology Studies (CWTS) (OpenAlex, 2024). Using a previously developed methodology, the CWTS group used just over 1.7 billion citation links to create 4,521 distinct clusters of publications (Van Eck & Waltman, 2024). These clusters were then labelled using the large language model (LLM) GPT 3.5 Turbo, developed by Open AI. Van Eck and Waltman (2024) provided the LLM with the 250 most-cited publications from each of the 4,521 clusters, and instructed it to create labels for each cluster based on the titles of the publications provided. The clusters and labels were then integrated into OpenAlex’s new structure as OpenAlex topics. The domains, fields, and subfields in OpenAlex are based on Scopus’s Subject Area Categories, Subject Area Classifications, and All Science Journal Classification (ASJC) Codes, respectively (OpenAlex, 2024). The addition of topics to OpenAlex effectively adds a fourth layer to the Scopus structure, which was developed by “in-house experts” (Scopus, 2024). The topics developed by CWTS were integrated into this existing structure; an LLM was used to assign topics to the most relevant subfield, and each topic was assigned to only one subfield. Because each subfield belongs to one field and one domain, each topic is automatically encompassed by those upper classes. This means that unlike in the poly-hierarchy that characterized OpenAlex’s early life, each class only has one parent class per level above it. Any field, subfield, or topic can only be reached by a single path, and as one descends down the hierarchy, the granularity of classes increases. To account for the works that were not assigned a topic during the development of the new hierarchy, and to allow for the incorporation of new publications into the OpenAlex structure (a crucial component of any research database), OurResearch fine-tuned a BERT deep-learning natural language processing (NLP) model (OpenAlex, 2024). When a work is added to OpenAlex, its title, abstract, citations, and journal name are used by this model to predict what topic is the best fit for this publication, and the work is then assigned to that topic and all its parent classes. At present, this process occurs automatically and can be understood as an instance of unsupervised machine learning. A final note on the structure of OpenAlex is that earlier in 2025, many of the topic labels were reworked following feedback from users (Demes, 2025). A quick check of the associated dataset reveals that 4,516 labels out of 4,521 were changed, and that in general, the new topic labels were less specialized than their predecessors (OpenAlex, 2025). Nothing about the structure of the topics was changed. Of relevance for this paper is the fact that all 107 topics covered under the Health Professions field were reworked; the topics covered in this report are therefore the recently re-labelled topics. In our analysis, however, if we find some topic labels 2 that seem particularly unsuited to the field, we may compare them with their predecessors for a better sense of if the issue is simply a matter of the topic’s name. Given that OpenAlex topics are so new, we have found little research that investigates their suitability for indexing, and no research that considers their suitability for individual subject fields. Salatino et al. (2024) highlight the heterogeneity of knowledge organization systems (KOSs) across disciplines, and highlight gaps related to the number of concepts and depth of concepts covered by various KOSs. They note that OpenAlex is one of just five KOSs they found that offer comprehensive coverage across subject fields. Drawing from Salatino et al.’s (2024) findings, Jenset et al. (2025) use a self-developed accuracy rating scheme to compare OpenAlex topics to the topics in a classification scheme they created, based largely off of OpenAlex’s methodology. A key distinction between their approach and OpenAlex’s for topic creation is that where OpenAlex only uses citation metrics to create their topic clusters, Jenset et al. (2025) also incorporate the semantic relatedness of the publications into the topic clustering process. The result of this dual approach is a classification system with over 29,000 topics, which they claim addresses key gaps related to coverage and granularity in the field of large-scale classification. Additionally, the authors use a supervised machine learning model to assign unclassified publications to the topics. In comparing their system to OpenAlex, they found that their topics labels were more consistently representative of their publications. However, several factors must be considered alongside these results. First, Jenset et al. (2025) pulled their sample of 33 million articles from Dimensions, citing too many issues with OpenAlex metadata quality to reliably use it in their study. Additionally, they limited their sample to articles and reviews. Given that they used the semantic relatedness of publications in their topic clustering, it is unclear if publications in languages other than English were included, and if so, how they were accounted for. It is also interesting that the authors only compare the suitability of their classification system to that of OpenAlex. Given the provenance of their data, one might assume that Dimensions would be a more appropriate choice for an initial comparative test, especially considering the multitude of document types in OpenAlex that were not incorporated into their classification scheme. Also of note is that despite the direct comparison between their classification system and OpenAlex, Jenset et al. (2025) do not appear to be trying to create an openly available product; in fact, nowhere in their paper do they state their intention to use this classification system in service of open science. One’s surprise at this fact may be attenuated by noting all three authors’ affiliations with oligopoly publisher Springer Nature. Two key points emerge from this analysis. The first is the need for an assessment of OpenAlex topics while accounting for the variety of publications in its repository. Factors such as document type or language may affect the suitability of a topic for a work, and a close evaluation of how individual publications are indexed within a single subject field may help highlight areas for improvement. The need to continue improving the existing OpenAlex structure leads to the second point, which is the value of fully open repositories for researchers. Even if Jenset et al. (2025) develop a database from their classification system that is superior in scope and granularity to anything presently available, it may not be open, which greatly 3 hinders the accessibility of the works it houses. Having a freely available, open source alternative is crucial in today’s scholarly ecosystem, and to this end, OpenAlex’s existing infrastructure should be carefully scrutinized so that it may be ameliorated. All this is not to say that the critiques levied against OpenAlex by Jenset et al. (2025) do not hold any weight. Their comments on insufficient metadata quality partially mirror findings from Culbert et al. (2025) who compare OpenAlex’s reference coverage to that of WoS and Scopus, and Céspedes et al. (2025), who specifically assessed metadata accuracy, coverage, and completeness in OpenAlex. However, these last authors also emphasized the strength of OpenAlex’s linguistic metadata, pointing to its value for aggregating scholarly outputs from across the world into a single open repository. Okamura (2023) also makes use of this global coverage, using data pulled from OpenAlex to identify international clusters of collaboration in research. At a smaller scale, Schares and Mierz (2023) use OpenAlex to assess a selection of publications by researchers at Iowa State University, once again demonstrating its value as an openly available tool for research. It is clear that OpenAlex has incredible potential for scholarly research, but as Priem et al. (2022) point out, it is still in its nascency. This paper aims to evaluate the structure and indexing of OpenAlex topics within the Health Professions field and its related subfields. Based on previous findings, particular attention will be paid to the suitability of topic labels, suitability of document metadata, and document language. 3 Evaluation 3.1 Evaluation of the classification system 3.1.1 Methods To evaluate the structure of the classification system, we first exported the amount of articles per topic. This allowed us to visualize the distribution of articles in the Health Professions field. We also visualized the distribution of topics per subfield. Each of the 107 topics were evaluated by a group member by comparing the title, the description, and the keywords. This comparison was also done for each of the 11 subfields. In comparing these elements, we answered the following questions with “Yes”, “No” or “Somewhat”. 1. Does the title properly describe the topic or subfield based on its description and keywords? 2. Do the descriptions and keywords make sense? 3. For topics, does this topic belong in the Health Professions field? “Somewhat” was selected when a title, description or keyword had a small error or if the answer to the question was partially true. 4 3.1.2 Results and Discussion In visualizing the amount of topics per subfield, it became quickly apparent that the topics were not evenly distributed. In fact the General Health Professions subfield has 50 topics which represents 46.7% of the topics in the Health Professions field. On the other hand, 4 subfields represent only 7.48% of the topics, with one of the subfields only having a singular topic. This evidently led to an uneven distribution of articles as shown in Figure 1. Figure 1: Distribution of Topics and Articles per Subfield This uneven distribution is further exacerbated by the poor descriptions of the subfields. For example, the description of the General Health Professions subfield that has an overabundance of topics is: “type of profession” (OpenAlex, n.d.). This is not nearly descriptive enough and provides no insights as to the type of topics that should be placed under this heading. The keywords are similarly ambiguous. Not only is the distribution uneven, but the structure of the subfields is unclear as well. For example, the Pharmacy subfield description is: “science that studies the chemical, pharmacological and pharmaceutical properties of drugs” (OpenAlex, n.d.). However, the 5 topics found under the Pharmacy subfield do not fit this description. Medical Malpractice and Liability Issues is about client complaints, legal issues and the quality of healthcare, Obesity and Health Practices is about the stigma surrounding weight and weight-loss, and Infant Health and Development relates to infant health with a focus on crying and sleeping issues. 5 Additionally, two topics should be moved out of the Health Professions field and under other fields in the Health Sciences domain. Oral and gingival health research, which explores risk factors surrounding oral health, should be moved to the Dentistry field, and Nursing care and research , which“focuses on the role of family caregivers” (OpenAlex, n.d.), should be moved to the Nursing field. These are the 5 topics that are under the Pharmacy subfield, but the titles, descriptions and keywords show no apparent reason for this structure. We found that 74.8% of the titles are appropriate descriptors for their topic. There are a variety of reasons a topic may not have been deemed to have an appropriate title. One issue encountered was topic titles that could be misleading if you did not know that they were in the Health Professions field. For example, there is a topic called Homelessness and Social Issues , but we believe it should be Homelessness and Health Issues as it is much more evocative of the articles that should be classified under this topic. Some topics have an unnecessary reference to a geographical location because of the naming mechanism discussed in Section 2. Given that the labels were created using the 250 most cited articles, this could mean that if a certain percentage of those articles were citing each other and referenced a geographical location, it can skew the data and determine incorrect keywords and themes, such as Social and Demographic Issues in Germany . This example has a reference to an unnecessary geographical location and is ambiguous unless you know the topic is under the Health Professions field. The last big issue was if the aboutness determined from the title was different from that determined by keywords and description. In other words, they were incorrect. An example of this is Trade Secret Protection Methods whose description is: This cluster of papers focuses on the intersection of healthcare, knowledge management, and information technology. It explores topics such as ethical considerations in healthcare, data mining for knowledge-driven healthcare processes, the integration of IT in healthcare delivery, and the protection of trade secrets in the context of healthcare innovation and management. The cluster also delves into the application of telehealth systems, e-health paradigms, and security systems for healthcare information. (OpenAlex, n.d.) It is also important to note that the titles for the topics were changed as mentioned in Section 2. The previous title was Knowledge Management in Healthcare which is much more descriptive and its aboutness matches that of the description, unlike Trade Secret Protection Methods . We would also like to highlight the incorrect change in title of Health Disparities in Roma Populations to Romani and Gypsy Studies . It is important to note that the titles were assessed solely by their aboutness. This means that the 9 titles that did not follow title case capitalization are not necessarily accounted for in the percentage of topics that are listed as “No” for the question “Does the title properly describe the topic or subfield based on its description and keywords?”. This is a small error to correct, but one that we would like to highlight here. 6 For the second question, it was deemed that 84.1% of the descriptions and keywords made sense and the other 15.9% were somewhat correct. This means that no descriptions or set of keywords were judged to be incorrect. The somewhat correct keywords and description often involved too much specificity. This could be by referencing geographical locations or specific groups such as the NHS. Lastly, given the interdisciplinary nature of health professions, some topics should have been placed in other fields. Out of the 11 total topics, which represents 10.3% of topics, that do not belong in the Health Professions field, 9 of them should be moved to fields within the Health Sciences domain. 4 should be moved to Dentistry , 3 should be moved to Nursing and one topic should be moved to each of the Veterinary and Medicine fields. The Romani and Gypsy Studies topic was also identified as inappropriate in the Health Professions field. Though it does discuss challenges in accessing healthcare, from the description and keywords it appears to be more about socioeconomic disparities and its effects on the Roma population. Doctoral Education Challenges and Solution is the last topic identified that does not belong in the Health Professions field. This topic references mental health issues faced by doctoral candidates, but it also references socialization, research supervision, publication rates and career trajectories. 3.2 Evaluation of indexing 3.2.1 Methods In order to assess how well articles are indexed under the Health Professions field, we performed an evaluation of a random sample of 80 articles, which spanned 25 topics, or about 23% of all the topics in the field. This was done by two group members. First, we both evaluated the same 10 articles to act as a baseline, and then we divided up the rest of the articles to evaluate individually. We assessed each article, determined its aboutness, and answered “Yes”, “No” or “Undetermined” to each of the following questions: 1. Is the topic assigned to the article relevant? 2. Is the topic assigned the best fit for the article? For articles where the answer to Question 2 was “No” or “Undetermined”, we attempted to assign a best fit topic. Then, for all articles, we answered the following: 3. Is the best fit topic a topic within the Health Professions field? A total of 30 out of the 80 articles assessed, or 37.5%, had titles in languages other than English, including French, Portuguese, German, and Indonesian. For any articles other than those in English and French, we translated the text using Google Translate. 7 When determining whether a topic was relevant or the best fit, we looked both at the title of the topic and its description, as well as other pertinent metadata if necessary. For Question 1, the answer “Undetermined” was assigned when there was not enough information about the article to determine its aboutness. If the title was vague, we looked to the abstract and other metadata. If that was missing or equally vague, we attempted to find the entire article, and if that was not available, then we assigned the “Undetermined” designation. For Questions 2 and 3, “Undetermined” was used both for the above scenario, and also if there was some ambiguity as to what the best fit topic should be, which will be further explained in the results section. 3.2.2 Results We found that of the 80 articles assessed in the Health Professions field, 65.5% of them were assigned to a relevant topic and 27.5% were not, while the remaining 7.5% (assigned “Undetermined”) did not have enough information for us to determine the aboutness of the article and therefore whether it was relevant (see Figure 2). This uncertainty carries through to the “Undetermined” assignations of the other parts of our analysis. Figure 2: Is the OpenAlex Topic relevant to the article? We looked at whether the language of the article titles had an impact on how well the articles were classified by OpenAlex. For articles with English titles, we found that 20% (10 out 8 Scopus. (n.d.). What are Scopus subject area categories and ASJC codes? Scopus Support Center. https://service.elsevier.com/app/answers/detail/a_id/12007/supporthub/scopus/ Van Eck, N.J., & Waltman, L. (2024, January 24). An open approach for classifying research publications. Leiden Madtrics . https://www.leidenmadtrics.nl/articles/an-open-approach-for-classifying-research-publicat ions Zafar, H. (2025). Examining the OpenAlex Concepts: A Detailed Case Study of Machine-Derived Classification [Master’s thesis, University of Ottawa]. uO Research. https://doi.org/10.20381/ruor-31060 15