This study was based on data obtained from PubMed [18] and Scopus [19] as of April 6, 2023. We chose Scopus over other databases because of its superior coverage in the health sciences field, accurate indexing, and “federated search interface” (ie, functionality), which enables us to query the content found across its sources using a common or standardized search form. The authoritative and comprehensive PubMed is the top choice for searching medical and health sciences literature. Medical Subject Headings (MeSH) terms were used for the PubMed search to improve comprehensiveness. Raw data were stored in TXT and CSV files, respectively. The R “Bibliometrix” tool of the R Statistical software [20] was used to clean data and integrate the 2 databases’ unique publications into a combined dataset (with and without assigned DOI) [18,19,21-23]. At this point, duplicate documents with an assigned DOI (1809 in total) were eliminated. The inclusion criteria for the outcome were all the research documents, written in English in peer-reviewed journals, that were published between 1974 (the earliest article we found) and 2023, and that dealt with RCC. Books, book chapters, gray literature, and reports were not included to avoid noise. The search strings, which used a Boolean computation, are indicated in Table 1. Only those articles with the term “renal cell carcinoma” in their title or abstract were selected. The leader and the other authors reviewed the complete list of all possible acceptable publications. Their reliability and value to the field were based on criteria like the journal’s impact factor, author affiliations, and citation count.

Our bibliometric review procedure adhered to best practice guidance published elsewhere [24] (Checklist 1). A preliminary descriptive analysis of the retrieved information was carried out using the R package bibliometrix [22]. This open-source application analyzes publication and citation metrics using mathematical and statistical methods to obtain a broad picture of the scientific output that was within the purview of the study. The questions answered were as follows: (Q1) What are the primary research topics in RCC? (Q2) How have RCC research questions changed over time? (Q3) How are these research topics distributed across countries and scientific journals? Three levels of analysis—countries, sources, and authors—were included at this stage to answer the above questions.

The annual growth rate of publications was calculated using the Bibliometrix package in R, which computes the compound annual growth rate (CAGR) using the following equation:

where V_f is the number of publications in the final year of the study period, V_i is the number of publications in the initial year of the study period, and n is the number of years between the initial and final year.

The unsupervised machine learning algorithm LDA [17] was applied to identify topics. Considered an extension of the probabilistic latent semantic analysis, it has its roots in Bayesian models [17,25]. Topics in LDA are thought of as multinomial distributions of vocabulary terms, in which each word has a given probability of occurring in a topic. This leads to the prominence of words that are more frequently used in a topic, creating clusters that reflect specific underlying themes. LDA does not require prior knowledge of the topics or the way they are presented in the texts. Rather, topics merely flow from the statistical properties of the data and the model’s underlying assumptions. For thematic analysis, it was decided to use abstracts rather than entire texts because topics are more coherent and ranked higher in large document collections. Inaccurate or noisy terms have less impact on topic word distribution [26].

The LDA model was validated using the Cv metric, which is grounded in the distributional hypothesis stating that words with similar meanings tend to coexist in similar contexts [27]. In other words, Cv rates the semantic similarity of words within a topic (ie, topic interpretability). The Cv score was calculated by looking at word co-occurrence statistics in a reference corpus and their conformity to human-like semantic interpretation. The package textmineR was used for such analysis, which made it easier to determine the ideal number of topics (k) for the study. Because higher scores indicate better interpretability, a model with the highest coherence score among those in the study rank (from k=4 to 50) was selected to reach a balance between granularity and thematic clarity.

The procedure for identifying topics through LDA was divided into 3 stages: (1) preprocessing, (2) construction of the LDA model, and (3) assigning labels to topics. LDAShiny [28], an open-source R package that uses Bayesian inference for LDA and machine learning algorithms to improve the analytical process, was selected for the first 2 stages.

For each topic, additional characteristics were revealed, especially at the journal and country levels, through statistical description based on the probability distributions of document-topic and topic-word acquired through LDA. To make results and findings more evident, we used certain quantitative indices suggested by Xiong and colleagues [33], which were obtained by adding document-topic and topic-word distributions. The indexes were described as follows.

The distribution of topics over time was obtained by the following equation:

where mϵj represents articles published each year, θ_mk is the proportion of the kth topic in each item, and n^y is the total number of articles published in the year.

Topic distribution across journals was defined as the ratio of the kth topic in the journal j: θkj as indicated in the following equation:

where mϵj represents the articles in a particular journal, θ_mk is the proportion of the kth topic on each item, and n^j is the total number of articles published in the journal j.

Topic distribution across countries was defined as the ratio of the kth topic in the country c, as in the following equation:

where mϵc represents the articles in a particular country, θ_mk is the proportion of the kth topic on each item, and n^c is the total number of articles published in the country c.

With the purpose of facilitating the characterization of the topics in terms of their tendency, topic datasets were also submitted to simple regression slopes where the year, country, and journal were the dependent variables, while the proportion of the topics in the corresponding year, country, and journal was the response variable [34]. From the regression slopes, we determined the directionality of these trends and set a significance threshold of P<.01 (Eq2). Topics that showed statistically significant positive slopes were identified as having upward trends, while those with statistically significant negative slopes were in decline. Tendencies were finally visualized using the ggcorrplot library of R to represent correlation strengths as a heatmap matrix. The color-mapped matrix was subjected to advanced hierarchical clustering analysis in order to investigate and compile correlation datasets, given its visual form like a tree-shaped dendrogram. The Agnes function with Ward’s method showed the agglomerative hierarchical clustering of variables. Each leaf of the dendrogram corresponded to one observation (variable), and the fusion height showed the dissimilarity between 2 observations on the vertical axis. A cut height for cluster identification was calculated using the Average Silhouette method [35].

The consolidated dataset was obtained by combining the results and removing duplicates, totaling 39,856 articles (Figure 1). After the relevant Excel file was created, 4628 articles lacking titles, abstracts, or affiliations were removed. Within this extensive dataset, 35,228 documents were assembled, demonstrating an annual growth rate of 9.86%. The extensive summary of the key descriptive characteristics pertaining to RCC from 1974 to 2023 can be found in Table 2. A substantial number of information sources, 3070 in total, were revealed by the data. Given that the document’s average age was 11.6 years (the time since the publication of the examined articles), it is likely that much of the research was carried out some time ago. On average, however, each document received 32.35 citations, demonstrating their influence and recognition in the field (Table 2).

With respect to the yearly output of documents on RCC, Figure 2 provides an overview of increased production from 1974 (24 articles) to 2023 (2401 articles). Throughout the 1980s and 1990s, the document production grew modestly. However, the 2000s saw a significant increase in production, reaching a peak in 2017 (1832 articles) and indicating a solid trend in recent years. Table 3 lists the top 30 scientific journals, while a global map (Figure 3) shows the 118 countries that were involved in RCC research. With 10,308 publications, the United States clearly was the top contributor.

The methodological rigor, reproducibility, and accuracy of LDA were ensured using advanced bibliometric tools such as Bibliometrix and Textminer. There were 30 topics with the best coherence score in the LDA model (Figure 4). The terms with the highest probabilities and semantically relevant labels for each latent topic are shown in Table 4.

The 30 identified topics were in 8 crucial domains of RCC research:

The topic distribution by document θm was added to compute the average probability θky of all the articles published in a particular year to identify the trends (Figure 5). We found that the probabilities of some topics steadily increased over time (red). Black indicates topics with no discernible trend, whereas blue denotes topics with a decreasing behavior.

Although the discovery of 30 unique themes with high coherence scores, a granular and nuanced analysis of the research landscape, was made possible by the LDA method, there was a need for validating the theme output by a form of visual data analytics. Heatmaps helped examine and understand how thematic patterns were related to variables like publication year, country, and journal. Red highlighting indicates the strongest associations among variables, reflecting higher correlation levels.

Figure 6A illustrates correlations between specific topics and years. Red highlighting indicates the strongest associations among variables, reflecting higher correlation levels. For example, within Cluster 4, Topic 16 (t_16) “Tumor Activity and Immune Responses” is primarily associated with the years 1993, 1985, 1988, 1990, 1992, 1991, 1989, 1995, 1994, 1996, and 1999. Topic 17 (t_17) “Nephrectomy and Recurrence” exhibits stronger correlations with the years 1993, 1990, 1992, 1991, 1989, 1995, and 1994. In Cluster 3, Topic 22 (t_22) “Gene Expression and Prognosis” significantly correlates with the years 2020, 2021, 2022, and 2023, while Topic 27 (t_27), also addressing “Gene Regulation and microRNA Expression,” shows significant associations with the years 2017, 2018, 2019, and 2020. Finally, in group 3, Topic Thirty30 (t_30) “Clinical Presentation and Diagnosis of Rare Cases” is linked with the years 1982, 1983, 1981, 1978, 1977, and 1975.

Figure 6B presents the interactions between countries and topics across different groups. In group 1, Topic 27 (t_27), “Gene Regulation and microRNA Expression,” is predominantly correlated with Taiwan. Group 2’s topics, “Tumor Activity and Immune Responses” and “Toxicity and Adverse Events in Treatments,” are closely associated with the Netherlands. For group 3, “Survival and Disease Progression” shows significant correlations with the Czech Republic, Denmark, Belgium, France, and Italy, and “Nephrectomy and Recurrence” is notably linked with Israel and South Korea. Lastly, in group 4, “Gene Expression and Prognosis” and “Gene Regulation and microRNA Expression” are primarily connected with China.

Figure 6C outlines the dominant publication patterns for specific topics within various journals. In group 1, “Tumor Activity and Immune Responses” is predominantly linked with the International Journal of Cancer and Cancer Research. Group 2’s topic, “Gene Expression and Prognosis,” has a significant association with Frontiers in Oncology. In group 3, “Survival and Disease Progression” relates to Clinical Genitourinary Cancer, and “Nephrectomy and Recurrence” is associated with BJU International, Journal of Endourology, and Urology. Lastly, in group 4, “Diagnostic Imaging and Radiological Evaluation” is prominently linked to the American Journal of Roentgenology. LDA algorithm’s identification of the key RCC research themes was mirrored in the relevant clustergrams that emerged from the hierarchical cluster analysis of the heatmaps.

By displaying each topic as a group of related words, LDA was able to identify latent (hidden) topics within a corpus of documents and demonstrate how each document may be represented as a combination of these topics. According to this method, RCC research has evolved over the past 50 years from concentrating on surgery to comprehending its genetic underpinnings and the impact of new treatments like immune checkpoint inhibitors and targeted therapies, which have improved the prognosis of metastatic RCC.

The disease characterization, progression, and pathological features have dominated the RCC research landscape for the past 50 years (t_16 and t_19) [10]. Kidney cancer was regarded as a single disease until the VHL gene was discovered [9]. Since then, scientists have realized that kidney cancer is a multitude of diverse diseases, each with its own genetic makeup. Although the characterization of genetic mutations and chromosomal alterations linked to the growth of RCC [36,37] has recently improved our knowledge of this kidney cancer (t_22), RCC research does emphasize the necessity of assessment of treatment effects’ safety and efficacy, as well as surveillance of RCC recurrence following nephrectomy. Therefore, a better understanding of resistance mechanisms, molecular prognosis (t_22), and immunological responses (t_16) is essential.

The advances in omics technologies over the last 10 years constitute a promising area of personalized RCC cure [1,38,39]. In addition, the microRNA signature in RCC and its function in progression, diagnosis, therapy targeting, and prognosis of RCC (t_27) have also received particular attention [40,41]. Recent RCC research also moves toward earlier cancer detection through broad imaging and radiological evolution (t-11). RCC has a difficult pathological classification, since the histological analysis reveals three most recurrent sporadic types: clear-cell RCC (70%‐75%), papillary RCC (10%‐15%), and chromophobe RCC (5%) [42]. It is predicated on morphologies, architecture, underlying genetic abnormalities, and tumoral protein expression [43,44]. Pathologists can detect these malignancies more accurately and provide better treatment plans and patient outcomes if they are aware of the significance of these markers (t_12 and t_19).

A major cause for worry is the postoperative surveillance of RCC recurrence following nephrectomy (t_17) [45]. Research on local RCC ablation and assessment of treatment effects on safety and efficacy is still ongoing (t_12, t_16, and t_19). The discovery of tailored medication like tyrosine kinase inhibitors (also called TKIs), such as Sunitinib and Sorafenib, has been beneficial in treating metastatic RCC [46,47]. Another well-established component of RCC treatment is checkpoint inhibitor immunotherapy [48], which shows its superior therapeutic efficacy when combined with TKIs [49]. The next generation of TKIs and immunotherapy (t_16) [50] is being developed to overcome some unfavorable outcomes with kinase inhibitors [51,52] and immunotherapy (t_19) [53], as well as the emergence of drug resistance [54]. Only a small number of themes still earn little attention. The topics “Serum Levels and Patient Control” (t_2), “Hypoxia Factors and Related Proteins” (t_7), and “Risk Prediction Models and Clinical Assessment” (t_18) are overlooked and should require additional attention.

In line with the LDA analysis and our own scientific expectations and goals, it was also feasible to identify untapped topics that have not yet been covered by scholarly literature. Intriguing lines of inquiry are the prognostic value of vascular endothelial growth factor [55], endostatin [56], C-reactive protein [57], the hypoxia-induced pathway [58], and ferroptosis [59]. The involvement of chronic inflammation [60] and gut and urinary microbiota in immune modulation of metastatic RCC [61] remains poorly investigated. Clinical judgments and patient stratification in RCC may be enhanced by the creation of new predictive models [62] based on, for instance, genetic biomarkers [63]. The application of artificial intelligence is another potential topic that could help physicians in identifying RCC subtypes by analyzing computed tomography scans, as well as in deconstructing complex epidemiological and environmental factors that influence RCC occurrence, like hypoxia [64].

In contrast to traditional bibliometric analysis, the LDA approach effectively extracted potential themes and inferred implicit information from a large collection of documents. The LDA approach and other topic modeling methods like co-citation and keyword co-occurrence cannot be compared with the same conceptual granularity or depth. LDA goes beyond who cites whom (ie, intellectual connection and research lineage) and uncovers the underlying conceptual themes that bind the literature, which may not be immediately apparent from citation patterns alone. Unlike a list of keywords, which reveals basic relationships, LDA organizes these co-occurring words into meaningful higher-order themes, offering a more detailed knowledge of topic relationships and structure. For this reason, the LDA analysis was not affected by the potential simplicity of the keywords chosen in the study.

The subjectivity involved in manually labeling LDA topics, the possibility of missing publications by using only 2 databases, the linguistic bias introduced by excluding articles written in languages other than English, the constraints of the “bag-of-words” model which disregards grammar and context, and the effects of excluding literature like book chapters are some limitations of the study. Despite the constraints, LDA can disclose “unknown unknowns” by revealing unarticulated or unacknowledged themes. It can also give an overview of the research landscape to identify new topics and interdisciplinary connections, as well as demonstrate how old themes are resurfacing in new ones. As a result, LDA remains the most often used natural language model [13,14,65].

This review offered a thorough summary of how research on RCC has changed over the previous 50 years. LDA helped identify important emerging trends in treatment development to address drug resistance and undesirable side effects, surgical techniques, and immunotherapy advancements, among other topics pertinent to clinical practice and medical research. In summary, this study presents a methodological synthesis of the development of RCC research and delivers pertinent data for clinical decision-making, early identification, and the planning of new biomedical research.