The Web of Science Core Collection (WoSCC) is a highly regarded database in the field of bibliometrics (11, 12), having gained significant recognition among scholars and serving as the primary data source informing this study. The search term strategy in this study was formulated by integrating reviews, bibliometric studies related to artificial intelligence and autoimmune diseases, as Medical Subject Headings (MeSH) terms in PubMed (5, 13). Supplementary Table S1 provides a list of the search terms employed in this study.

The search was conducted on July 1, 2024, with the search terms limited to English language articles and reviews published between 2003 and the search date. After the retrieval and screening process, a total of 1,695 publications were identified, comprising 1,409 research articles and 286 reviews. The WoSCC was used to systematically collect data on publication countries/regions, institutions, journals, authors, articles, and keywords. Excel software (version 2021) was used to conduct summary and analysis.

CiteSpace is a Java application created by the research team led by Chen Chaomei at Drexel University (14). This software offers visualization techniques for publication data and has been widely acknowledged in the field of bibliometrics for providing objective analyses of academic frontiers (15, 16). In the CiteSpace analysis, co-occurrence networks were used to distinguish merged networks via color-coded nodes and edges. Burst detection, based on Kleinberg’s algorithm, was employed as an indicator of active topics (17). In this study, CiteSpace charted the literature network within the relevant field, offering analyses on research hotspots, frontier trends, and the evolution of knowledge structures.

VOSviewer, developed by Leiden University in the Netherlands, was utilized to extract and analyze elements within literature data such as authors, keywords, and institutions (18–20). It calculated the strength of their associations and presented them in a visual format. This process involved data extraction and preprocessing, association calculation, and visualization mapping to facilitate co-occurrence analysis, citation analysis, and clustering analysis. VOSviewer was utilized to create a node network, and relevant information was obtained through parameters such as link strength, cluster color, and node size analysis.

HistCite Pro 2.1 was employed to manage and analyze a large volume of literature data, excelling particularly in citation analysis (21). By utilizing HistCite, citation relationships between documents could be traced, elucidating the knowledge dissemination pathways and developmental trends within the research domain, thereby identifying highly significant literature in the field. In HistCite Pro 2.1, the “Limit” was set to 30, with the remaining settings kept at their default values. Subsequently, the “Make graph” option was selected to effectively and visually represent the interconnectedness within the research field, facilitating the efficient identification of key literature.

Bibliometric analysis was able to quantitatively portray the developmental status within a specific academic research field. Our study revealed that over the past two decades, the total number of publications in this research field amounted to 1,695. Among these, 1,409 were research articles and 286 were reviews. These publications include the contributions of 10,915 authors from 7,070 institutions across 703 journals ( Supplementary Table S2 ).

As depicted in Figure 2A , we observed a rising trend in the number of publications over the past two decades, with particularly accelerated growth in the last five years. The count surpassed 100 for the first time in 2020, reaching 139 publications. By 2023, it had reached 341 publications, and in the first half of 2024 alone, the count had already reached 222 publications. In terms of citations, there was a noticeable upward trend in the number of citations in 2017, surpassing 1,000 and reaching 1,152. It is noteworthy that by 2023, the number of citations had reached 5,777.

A thorough examination revealed that over the past two decades, publications on AI in AID spanned 124 Web of Science Categories. As illustrated in Figure 2B , the top 20 Categories with the highest publication counts are showcased. Among these, there are 5 Categories with over 100 publications, with Rheumatology leading at 222 publications, followed by Immunology with 189 publications. Subsequently, there were 117 publications in Medicine General Internal, 110 publications in Multidisciplinary Sciences, and 105 publications in Surgery.

In the field of research, a total of 57 countries/regions contributed publications, with the USA having the highest number of publications at 506, accounting for 21% of the total, followed by China with 416 publications, making up 17%. Additionally, Germany had 137 publications, while England and Italy each had 136 publications. The pie chart in Figure 2C illustrates the distribution of publications by different countries/regions, revealing that the top 12 countries accounted for 73% of the total publications. The global distribution of publications by countries/regions, as depicted in Figure 2D , overall indicates that the United States and China made the most significant contributions to the field in terms of publication output. Unfortunately, it is noted that countries in Central Asia and regions in Africa have not yet ventured into this research field.

Through the analysis of co-authorship relationships among countries/regions using VOSviewer, it was found that the USA had the highest total link strength, reaching 437, indicating a higher frequency and wider scope of collaboration with other countries ( Figure 3A ). In terms of publication citations, a minimum citation threshold of no less than 5 times was set, and a country/region citation network map was generated ( Figure 3B ). In this network, node size represents citation counts, with the top five countries in citation counts being the USA, Germany, China, England, and Italy, with citation frequencies of 13,515, 3,425, 3,298, 3,277, and 2,436 respectively. Furthermore, through the analysis of total link strength, the USA, Italy, and China ranked in the top three with values of 1,849, 1,040, and 1,023 respectively, indicating the wider dissemination of publication citations from these three countries/regions.

Various institutions leverage their expertise in clinical medicine, data algorithms, and biological research to collaboratively consolidate resources and knowledge. Setting the publication threshold at a minimum of 8 articles, a collaboration network comprising 88 academic institutions was established ( Figure 3C ). Node size represented publication output, with Brigham & Women’s Hospital, Harvard Medical School, Sichuan University, Mayo Clinic, and University of California San Francisco ranking as the top five institutions, with publication counts of 37, 36, 25, 23, and 23 respectively. Moreover, in this network, Brigham & Women’s Hospital, Harvard Medical School, and Massachusetts General Hospital ranked among the top three for total link strength, with values of 91, 67, and 60 consecutively.

By focusing on citations, with a minimum threshold of 100 citations, a total of 223 academic institutions were identified in the network graph ( Figure 3D ). The analysis revealed that Harvard University claimed the top position with 1,567 citations, followed by Brigham & Women’s Hospital with 1,372 citations and Massachusetts General Hospital with 1,129 citations. In the citation network graph, Brigham & Women’s Hospital was identified as the institution with the highest total link strength, with a value of 743. Massachusetts General Hospital and Harvard University recorded 638 and 597, respectively. These findings indicate that these three institutions occupy a dominant position within the field.

Interdisciplinary collaboration among authors overcomes professional barriers and drives progress in the field. Based on a minimum threshold of 3 publications, an analysis was performed to construct a co-authorship network graph for 323 authors ( Figure 4A ). The analysis revealed the existence of distinct co-author clusters, with notable prominence observed in clusters spearheaded by Gainer, Vivian S. and Karlson, Elizabeth W., Cai, Tianxi and Liao, Katherine P., as well as Kleyer, Arnd, Simon, David, and Schett, Georg. These findings suggest that the aforementioned individuals demonstrated higher levels of involvement and collaboration in research activities within their respective clusters. With regard to the presentation of citations, Figure 4B depicts the citation network graph, wherein the nodes represent the citation counts. Cai, Tianxi was the most highly cited author, with 955 citations, and also ranked first in total link strength within the network, with a value of 106. A co-citation network graph was constructed based on a minimum citation frequency threshold of 20 times, comprising a total of 219 authors ( Figure 4C ). In this co-citation network, the highest number of citations were attributed to Smolen, JS; Breiman, L.; and Aletaha, D., with respective citation counts of 177, 136, and 129.

We also analyzed the total link strength and norm. Citations of the top 15 authors in terms of publication volume ( Figure 4D ). Our results revealed that the author with the highest publication volume was Cai, Tianxi, with 17 articles, and the highest total link strength of 106. However, Cai, Tianxi ranked second in the norm. Citations with 22.10. The author ranked first in the norm. The citation was Liao, Katherine P., with a publication volume of 12 and a total link strength of 92, which placed them second. This indicates the significant influence of Cai, Tianxi, and Liao, Katherine P. in the field of study.

Through an analysis utilizing VOSviewer of journal citation networks, it was observed that Autoimmunity Reviews, Annals of Thoracic Surgery, European Journal of Cardio-Thoracic Surgery, Scientific Reports, and Lupus Science & Medicine emerged as the top five journals based on total link strength, achieving scores of 134, 114, 111, 105, and 99 respectively ( Figure 5A ). These findings suggest a pronounced citation relationship with other journals. Moreover, an examination of the temporal aspect of journal citations, as illustrated in Figure 5B , revealed that yellow nodes denote journals that have displayed heightened activity in the field in recent years, whereas progressively bluer nodes signify earlier periods of relative activity for these journals. Furthermore, by setting the minimum citation threshold to no fewer than 100, a total of 154 core journals were utilized to construct the co-citation network ( Figure 5C ). The results indicated that the total link strengths for Annals of the Rheumatic Diseases, Nature, Proceedings of the National Academy of Sciences of the United States of America, PLOS One, and Journal of Immunology were 88,303, 72,645, 60,534, 58,811, and 57,919, respectively.

In the realm of academic publication output, Frontiers in Immunology emerged as the frontrunner with 86 articles, closely trailed by Scientific Reports with 50 articles, Rheumatology with 35 articles, Arthritis Research & Therapy with 31 articles, and PLOS One with 25 articles. Additionally, in terms of scholarly impact measured by citations, the Journal of the American Medical Informatics Association claimed the top position with 808 citations, followed by Scientific Reports with 745 citations, Frontiers in Immunology with 732 citations, PLOS One with 681 citations, and Annals of Thoracic Surgery with 660 citations ( Figures 5D, E ).

The dual-map overlay function illustrated the distribution of citing and cited journals, revealing the interdisciplinary connections between the research field and other disciplines. In Figure 5F , the dual-map overlay showed citing journals on the left side and cited journals on the right side. Two significant citation pathways were observed. The yellow path indicated that journals in the fields of molecular biology and immunology tended to cite publications from molecular biology, genetics, as well as health, nursing, and medical disciplines. The green path, on the other hand, suggested that medicine, medical, and clinical journals were inclined to cite publications from molecular, biology, genetics, as well as health, nursing, and medical fields.

In this study, we observe an increasing trend in the number of publications in the field over the past twenty years, with a particularly rapid growth in the past five years. The number of publications surpassed 100 for the first time in 2020, reaching 139, and by 2023 had escalated to 341. In just the first half of 2024, the number had reached 222. This growth trend indicates a continuous rise in research interest in the field, possibly influenced by factors including advancements in related technologies, increased societal attention, and heightened research investments. However, we must also acknowledge that this rapid growth trend may bring about certain challenges. With the increasing number of publications, there could be variations in research quality, necessitating more rigorous peer review mechanisms to ensure the reliability and effectiveness of research.

Both the United States and China have made substantial contributions in this field, however, there is an evident developmental imbalance among countries and regions worldwide. Patients with AID in these two countries benefit from more accurate diagnosis and effective treatment, ultimately improving their quality of life and reducing the suffering and societal burden imposed by these illnesses. Publications and collaborations in the field from African and Central Asian countries are relatively scarce, potentially due to limited healthcare resources and research funding in these regions, leading to disparities in the application of artificial intelligence in medicine. Further analysis supports the notion that countries with fewer publications or citations are predominantly from middle- to low-income or low-income areas.

Although automated, AI-driven medical applications can overcome barriers to healthcare equity in middle- to low-income countries, AI medical models trained in high-income regions may exhibit suboptimal performance when applied in middle- to low-income countries due to the phenomenon of “data set shift,” where differences in patient populations, clinical practices, and healthcare systems lead to discrepancies (48, 49). Hence, it is essential for the future to enhance collaboration with middle- to low-income countries to ensure model generalizability by collecting data from diverse and representative populations across various countries and regions, thereby preventing data set shifting (50).

Through the analysis of key terms in the research field, the main applications of AI in autoimmune diseases have been identified. AI is most commonly utilized in diseases such as multiple sclerosis, rheumatoid arthritis, and systemic lupus erythematosus, possibly due to the high prevalence and significant clinical attention these diseases receive. The application scenarios of AI in this field encompass six main areas: patient identification, risk factors and prognosis assessment, diagnosis, classification of disease subtypes, monitoring and decision support, and drug discovery ( Figure 10 ) (5, 51, 52).

AI shows potential in the field of autoimmune diseases, but it faces numerous problems and challenges. In terms of models, there is insufficient validation, necessitating enhanced cross-validation and independent testing to improve accuracy and reliability. Complex models are in demand, yet they encounter difficulties in computational power, methods, and data management. Data integration from multiple sources is challenging and requires standardization. Additionally, there is a lack of comprehensive “healthy” immune response models to differentiate between normal and diseased states. In treatment, the testing of combination therapies is hindered by numerous possibilities, requiring improved computational methods to assess their effectiveness and safety, while regulatory acceptance of digital evidence remains uncertain. Autoimmune diseases are complex and heterogeneous, posing difficulties in model construction and precise treatment implementation. Understanding disease mechanisms and developing advanced algorithms are crucial to address these challenges and drive progress in the field. Furthermore, there are significant privacy and security risks in medical health data, making the protection of patient data privacy a critical consideration when utilizing AI for autoimmune disease data processing (3, 5, 133–137). To address these challenges, future research should enhance cross-validation and independent testing, improve computational power, optimize data management and methods, promote data standardization, and develop universal and practical models. Furthermore, researchers must prioritize data privacy protection to ensure legal and compliant usage. Interdisciplinary collaboration should follow ethical guidelines to safeguard patients from potential harm and support the safe and reliable advancement of the field.

AI shows significant potential in the field of autoimmune diseases. The upcoming directions encompass enhanced diagnostic precision and subcategorization achieved by amalgamating diverse datasets and refining deep learning algorithms (138, 139). In the field of personalized therapy, AI is poised to expedite drug discovery processes and investigate combined treatment approaches (140). To facilitate disease surveillance and prognosis evaluation, AI can use wearable technologies and construct predictive models for continuous monitoring and risk assessment (105, 141). Furthermore, fostering interdisciplinary partnerships between medical and engineering sectors is imperative, particularly in the innovation of novel technologies such as digital twins (142). The establishment of extensive databases through data exchange and collaborative research initiatives across institutions will play a pivotal role in advancing autoimmune disease management (5).

It should be noted that there are certain limitations to bibliometric analysis in practical implementation. Due to variations in inclusion criteria and sources among different databases, the publications covered may differ. However, mainstream analysis tools currently struggle to fully eliminate the impact of these differences. Therefore, the choice of using WoSCC as the primary database aims to ensure the credibility and reliability of the data to the greatest extent possible. Additionally, focusing solely on English literature may lead to some omissions, and alterations in institution names could also introduce biases. In summary, we acknowledge these potential issues in conducting bibliometric analysis and have referenced high-quality bibliometric literature in the research process to enhance the accuracy and reliability of the analysis results.