Introduction
The coronavirus disease 2019 (COVID-19) pandemic profoundly affected global health systems and scientific research, disrupting clinical practice and healthcare delivery while altering research priorities [
1,
2]. In oncology, the pandemic disrupted cancer screening, diagnosis, and treatment and simultaneously increased attention to healthcare-system resilience and public health preparedness [
3–
5]. These changes may also have influenced the thematic direction of research in cancer epidemiology and prevention, which contributes to reducing the cancer burden through risk factor modification, screening, early detection, and population-based interventions [
6–
8].
Advances in natural language processing and machine learning have enabled large-scale analyses of the scientific literature. Topic modeling is increasingly used to identify latent research themes and track their evolution across biomedical disciplines. Previous studies have applied these methods to examine publication trends in oncology, COVID-19 research, and primary health care, demonstrating the utility of computational text mining in bibliometric research [
9–
14]. However, few studies have examined whether research topics in cancer epidemiology and prevention changed systematically during the COVID-19 era or whether the patterns of change differed among journals with distinct editorial scopes.
Unlike previous bibliometric studies that primarily characterized publication patterns or thematic structures, the present study combined topic modeling with inferential statistical analyses to quantify temporal changes in research themes. By comparing 3 leading journals covering cancer epidemiology, prevention, and related areas of oncology, the study also evaluated whether temporal changes in topic prevalence differed among journals. This approach was intended to provide a more detailed assessment of changes in scientific priorities across journals with differing disciplinary scopes.
Accordingly, we analyzed articles published from 2016 to 2025 in the British Journal of Cancer (BJC), Cancer Epidemiology, Biomarkers & Prevention (CEBP), and European Journal of Cancer Prevention (EJCP). Using latent Dirichlet allocation (LDA) topic modeling together with complementary statistical analyses, we compared topic distributions between the pre-pandemic and post-pandemic-onset study periods and evaluated whether temporal changes differed among journals. We hypothesized that (1) the distribution of research topics would differ significantly between the 2 periods and (2) the magnitude of these temporal changes would vary according to journal scope.
Materials and Methods
To evaluate temporal changes in cancer research topics, we analyzed articles published from 2016 to 2025 in 3 representative cancer journals. Bibliographic data were retrieved from PubMed through the publicly available E-utilities application programming interface provided by the National Center for Biotechnology Information (NCBI), which enables programmatic access to indexed biomedical literature without requiring subscription-based access [
15].
Articles published from January 1, 2016, through December 31, 2025, were identified using the following query:
"{Journal Name}"[Journal] AND ("2016"[pdat]: "2025"[pdat]).
For each journal, PubMed identifiers (PMIDs) were first retrieved using the esearch utility, after which complete bibliographic records were obtained in XML format using efetch. Records were downloaded in batches of 200 articles, with a 0.4-second delay between requests, in accordance with the NCBI E-utilities usage guidelines [
15].
Journal Selection and Study Scope
Three journals were selected to represent complementary domains of cancer research: BJC, CEBP, and EJCP.
These journals were selected for 3 reasons. First, their scopes encompass distinct components of the cancer research continuum, including clinical oncology (BJC), molecular and epidemiological prevention research (CEBP), and population-based cancer prevention (EJCP) [
6]. Second, all 3 are long-established, peer-reviewed journals indexed in PubMed, providing consistent bibliographic metadata and high-quality indexing [
15]. Third, their complementary editorial scopes enable comparison of thematic changes across clinical, epidemiological, and preventive cancer research [
6].
After records from the 3 journals were combined and duplicate entries were removed, the final dataset comprised 8,694 articles.
Data Extraction and Preprocessing
The following variables were extracted from each article: publication year and month, journal volume and issue, article title, author names, digital object identifier, and PMID. XML records were parsed using Python’s xml.etree.ElementTree module. Text from nested XML title tags was concatenated to preserve complete article titles, and author names were standardized to the “LastName ForeName” format.
Only article titles were used for topic modeling because they provide concise, standardized summaries of each publication’s primary research focus and therefore support consistent comparisons across journals and publication years. Titles are also less affected than abstracts by differences in manuscript structure, reporting style, and journal-specific formatting, making them suitable for analyses of broad thematic trends.
Before topic modeling, article titles were processed using a standardized text-cleaning pipeline [
16]. All titles were converted to lowercase, and punctuation, numerical characters, and special symbols were removed. English stop words were excluded using the built-in stop-word list in scikit-learn. The cleaned corpus was then transformed into a term-frequency document-term matrix using CountVectorizer, which served as the input for LDA. This preprocessing pipeline was applied uniformly to all articles.
The number of topics was determined through an iterative exploratory process in which multiple candidate models were evaluated. The final 10-topic model was selected for its interpretability, semantic coherence, and limited conceptual overlap among topics. This approach is consistent with previous applied bibliometric studies that prioritize interpretability and semantic coherence when selecting topic models.
After model estimation, topics were interpreted by jointly examining the highest-probability keywords and representative article titles assigned to each topic. Topic labels were manually assigned to reflect the dominant conceptual themes. Topic validity was assessed based on keyword coherence and consistency with representative publications. Because the primary objective was to characterize broad thematic patterns in cancer research, interpretability and conceptual distinctiveness were prioritized over quantitative model optimization metrics.
Computational Environment
All computational analyses were performed in Python ver. 3.12 (Python Software Foundation). Bibliographic records were retrieved using the requests library and processed with pandas. Text preprocessing and LDA topic modeling were implemented using scikit-learn; the Pearson chi-square test was performed using SciPy; logistic regression analyses were conducted using statsmodels; and analytical outputs were exported to Microsoft Excel 2024 (Microsoft Corp.) using openpyxl. The entire workflow—including bibliographic retrieval, preprocessing, topic modeling, statistical analysis, and data export—was implemented in Python to support analytical reproducibility.
Topic Modeling
Latent thematic structures were identified using LDA implemented in scikit-learn [
17]. Based on the preprocessed document-term matrix described above, candidate models with different numbers of topics were evaluated during model development. The final 10-topic model was selected because it yielded a semantically coherent, conceptually distinct, and substantively interpretable topic structure while maintaining stable topic assignments across repeated model fitting, consistent with previous applied topic modeling studies [
18,
19]. Each article was then assigned to the topic with the highest posterior probability.
Because the primary objective was to evaluate pandemic-era changes in cancer epidemiology and prevention research, Topic 7 was specified a priori as the representative topic for logistic regression analysis. This topic was selected based on its semantic coherence, conceptual interpretability, and close alignment with the study hypothesis, as reflected in its representative keywords and article titles. Although quantitative measures such as coherence scores and perplexity are commonly reported in methodological studies, they were not used as the sole criteria for model selection because such measures do not necessarily identify the most substantively meaningful topic structure in applied biomedical topic modeling.
Definition of Study Periods
To evaluate temporal changes in research trends during the COVID-19 era, articles were classified into 2 periods: (1) pre-pandemic period: 2016–2019; (2) post-pandemic-onset period: 2020–2025.
A binary variable (PostCOVID) was created and coded as 1 for articles published in 2020 or later and 0 for articles published before 2020.
Statistical Testing
Chi-square test
Overall differences in topic distributions between the 2016–2019 and 2020–2025 periods were evaluated using the Pearson chi-square test of independence applied to a contingency table of topic frequencies [
20].
Logistic regression analysis
To quantify the temporal change in the predefined representative topic, logistic regression was performed with topic membership as the dependent variable (1=article assigned to Topic 7; 0=otherwise) and study period as the independent variable [
21]:
logit(P(Topic))=β0+β1(PostCOVID).
Interaction model across journals
To examine whether temporal changes in Topic 7 differed among journals, the following interaction model was estimated [
21]:
logit(P(Topic))=β0+β1(PostCOVID)+β2(Journal)+β3(PostCOVID×Journal).
BJC served as the reference category. Interaction terms were used to test whether changes in Topic 7 prevalence between 2016–2019 and 2020–2025 differed significantly between each comparison journal and the reference journal.
Statistical Significance
The logistic regression analysis was intended to quantify temporal changes in the predefined representative topic, Topic 7, rather than changes across all extracted topics. Overall differences in the distribution of all identified topics were evaluated using the Pearson chi-square test. All statistical tests were 2-sided, and
p<0.05 was considered statistically significant [
20].
Results
Topic Modeling Results
A total of 8,694 articles published from 2016 to 2025 across the 3 selected cancer-related journals were included in the analysis. Topic modeling identified 10 distinct thematic clusters (Topics 0–9), each representing a coherent domain of cancer epidemiology, prevention, or clinical research.
The 10 topics broadly represented 3 domains. Primary prevention topics included lifestyle-related risk factors, environmental and occupational exposures, and preventive interventions. Secondary prevention topics included cancer screening and risk prediction based on biomarkers. Finally, non-prevention topics included clinical oncology, genetic and molecular epidemiology, epidemiological trends, and health disparities.
Articles were distributed relatively evenly across the thematic areas, with no single topic dominating the corpus. However, visual inspection of topic prevalence suggested modest shifts between 2016–2019 and 2020–2025 in the relative prominence of several topics, particularly those related to lifestyle risk factors, preventive strategies, and broader public health themes.
To facilitate interpretation of the topic modeling results, each of the 10 topics was assigned a descriptive label based on its most salient keywords and overall thematic coherence. As shown in
Table 1, the topics span a broad range of research domains, including cancer screening and early detection, molecular biomarkers and genetics, lifestyle-related risk factors, clinical treatment, and infection-related cancers. Several topics also reflect broader public health themes, including environmental exposures and prevention policy. This classification provides a framework for interpreting the subsequent temporal analyses and comparing changes across specific research domains.
Statistical Analysis of Topic Distribution
Changes in topic distribution before and after the onset of the COVID-19 pandemic
The relative distribution of research topics was compared between the pre-pandemic period (2016–2019) and the 2020–2025 period using the Pearson chi-square test (
Table 2).
The relative topic composition differed significantly between the 2 periods (χ²=21.42, p=0.011). Although most topic proportions remained relatively stable, modest increases were observed in Topic 2 (lifestyle risk factors) and Topic 7 (immunology, public health, and COVID-19–related oncology), whereas Topic 4 (epidemiological methods and cohort studies) and Topic 6 (infection-related cancers) showed relative declines during 2020–2025.
To further quantify the temporal change in a representative topic, logistic regression was performed for Topic 7. Topic 7 was selected for regression because it was interpretable and closely aligned with pandemic-era research themes (
Table 3). This hypothesis-driven selection was made before inferential testing and was based on conceptual relevance rather than the statistical significance of the observed association. Although Topic 6 more directly represented virus-associated carcinogenesis, Topic 7 was selected because it encompassed broader immunological, inflammatory, and public health themes considered more closely aligned with the wider research context of the COVID-19 era. The odds of an article being classified as Topic 7 were significantly higher during the 2020–2025 period (β=0.168, standard error [SE]=0.068, z=2.472,
p=0.013). This coefficient corresponds to an odds ratio of approximately 1.18, indicating a modest increase in the odds of Topic 7 classification relative to the pre-pandemic period. Topic 7 primarily represented themes related to immunological mechanisms, inflammation, and pandemic-era public health considerations. Its increased representation during 2020–2025 may reflect greater research attention to the intersection of immunological mechanisms, inflammation, and public health issues during the COVID-19 era.
Differences Across Journals
To assess whether the temporal change in Topic 7 differed among journals, an interaction model including journal and study period was estimated (
Table 4). With BJC as the reference category, the main effect for the 2020–2025 period was positive and statistically significant (β=0.245, SE=0.075, z=3.264,
p=0.001), indicating increased odds of Topic 7 classification in BJC during 2020–2025 relative to 2016–2019.
The interaction term for CEBP was negative and statistically significant (β=−0.959, SE=0.261, z=−3.669, p<0.001), indicating that the temporal change in Topic 7 differed significantly from that observed in BJC. In contrast, the interaction term for EJCP was not statistically significant (β=−0.383, SE=0.347, z=−1.105, p=0.269), providing no evidence that its temporal change differed from that in BJC.
Overall, these findings indicate that while topic prevalence shifted significantly after the onset of the COVID-19 pandemic, changes were not uniform across journals and varied substantially depending on journal scope.
Discussion
This bibliometric analysis showed that research topics in cancer epidemiology and prevention changed during the COVID-19 era, although the magnitude of these changes was modest and varied across journals. Topic modeling combined with inferential statistical analyses identified a significant difference in the overall distribution of research topics between the pre-2020 and 2020–2025 periods. However, the small effect sizes and limited explanatory power of the regression model indicate that the observed pattern is more consistent with modest thematic reorientation than with a fundamental restructuring of cancer research agendas. Previous bibliometric and meta-research studies have similarly reported incremental but statistically detectable changes in scientific output following major global events [
18]. Overall, the findings partially support the study hypotheses by demonstrating temporal changes in topic distribution while also showing that these changes differed across journals [
19].
The increased odds of Topic 7 classification (odds ratio≈1.18) indicate modestly greater representation of themes related to immunology, public health, and COVID-19–related oncology during 2020–2025. This pattern is consistent with previous studies reporting increased scientific attention to infection-related diseases, immune mechanisms, and health-system resilience during the COVID-19 era [
20–
22]. Nevertheless, the modest odds ratio and very limited explanatory power of the logistic regression model (pseudo R²≈0.001) warrant cautious interpretation. Rather than indicating a major transformation of cancer research, the result is more appropriately interpreted as a small but detectable difference in thematic emphasis within the large bibliographic dataset analyzed [
23].
Another notable finding was that the temporal pattern for Topic 7 was not consistent across journals. Interaction analysis showed that the 2020–2025 change in CEBP differed significantly from that in BJC, whereas the corresponding change in EJCP did not differ significantly from the reference journal. These findings are consistent with the possibility that editorial scope and disciplinary orientation shape how research communities respond to major public health events. Previous studies have similarly reported that journal-specific editorial priorities and target audiences may influence the adoption and dissemination of emerging research topics [
24–
26].
This study extends previous applications of topic modeling in biomedical research. Prior studies have used machine learning approaches to examine thematic changes in gynecologic oncology, renal cell carcinoma, COVID-19 research, and primary health care [
11–
14]. Whereas these studies primarily characterized thematic structures and publication trends, the present study combined topic modeling with inferential statistical analyses to quantify temporal differences before and during the COVID-19 era and to compare the Topic 7 temporal pattern across journals. This integrated approach illustrates how computational text mining can be combined with conventional statistical methods to evaluate temporal patterns in bibliometric data.
The findings also have implications for understanding how scientific research priorities change during major public health events. The observed temporal differences suggest that even highly specialized fields such as oncology may undergo measurable shifts in research emphasis during global health emergencies [
10,
27]. At the same time, heterogeneity in Topic 7 across journals indicates that temporal patterns may reflect not only broader external events but also differences in editorial priorities, disciplinary traditions, and institutional contexts that shape the production and dissemination of scientific knowledge [
28,
29].
Several limitations should be acknowledged. First, topic modeling was based exclusively on article titles rather than abstracts or full texts. Although titles provide concise, standardized summaries of research focus and are commonly used in large-scale bibliometric analyses, they contain less semantic information than abstracts and may therefore not fully capture finer thematic distinctions. Future studies incorporating both titles and abstracts may provide greater thematic resolution. Second, topic labeling remained subjective despite efforts to ensure semantic and conceptual consistency. In addition, assigning each article to a single dominant topic may not fully capture overlap among research themes. Third, the analysis was limited to 3 selected journals and therefore may not represent the full diversity of global cancer research. Finally, because this was an observational bibliometric study, the identified temporal differences should not be interpreted as evidence that the COVID-19 pandemic caused the observed changes [
30].
Despite these limitations, this study makes several contributions. Unlike previous topic-modeling studies that primarily characterized thematic structures within individual journals or research fields [
11–
14], this study combined computational topic modeling with inferential statistical analyses to compare temporal patterns before and during the COVID-19 era and to assess journal-specific differences in Topic 7. This approach provides a reproducible framework for examining temporal patterns in scientific research and illustrates the potential value of combining machine learning methods with conventional statistical analyses in bibliometric research [
31].
