Artificial intelligence in primary health care: A bibliometric analysis of publications from 2015 to 2024

Abstract

Objective

Primary health care (PHC) is a key approach to achieving the goal of “Health for All,” and artificial intelligence (AI) can empower primary health care in various contexts, including screening, diagnosis, and treatment. Since 2019, the application of AI in primary health care has attracted increasing academic attention. This is a bibliometric study aiming to identify publication trend, distributions, frontiers, and hotspots of extant research about AI in PHC.

Methods

We retrieved papers from the Web of Science Core Collection. Using VOSviewer, CiteSpace, and Bibliometrix, we conducted a bibliometric analysis to examine the publication trend, citation trend, country distribution, institution distribution, journal distribution, author distribution, reference distribution, keyword co-occurrence, and keyword burst.

Results

We totally obtained 653 English-language papers published in Web of Science Core Collection from 2015 to 2024. These papers were authored by 4304 researchers in 1551 institutions from 70 countries. Both publication volume and citation frequency have increased rapidly since 2019. The USA and the United Kingdom have the most publications, citations, and collaborations in this field. The journals that publish papers on this topic are mainly from the medicine field, spanning a broader range of disciplines, including molecular biology, genetics, health, nursing, medicine, psychology, education, and the social sciences.

Conclusions

This research area has attracted growing academic attention since 2019, and this trend is ongoing. Key research topics include diagnosis, family medicine, mental health, medical image analysis, and early screening of diseases.

Introduction

Primary health care (PHC) is the basic health care that is universally accessible to individuals and families in the community through their full participation and at a cost that the community or country can afford. ¹ It is the initial point of contact for the majority of people seeking medical care. ² Before the 1970s, various approaches and ideas failed to establish a well-functioning health care system capable of delivering equitable and high-quality care for people in the global range. ³ Therefore, WHO and UNICEF jointly organized a conference in Alma-Ata in 1978, where the goal of achieving “Health for All” by 2000 was introduced. ⁴ Further, the Alma-Ata conference proposed the concept of PHC, and pointed out that it is an effective way to achieve the “Health for All” goal. ⁴ PHC primarily includes services provided by general practitioners and family doctors at the community level, including health education, health management, disease prevention, early screening, general medical care, and community rehabilitation. ⁵ These services advance the goal of “health for all” through two mechanisms: reducing costs and improving health outcomes. First, PHC allows individuals to access diagnostic and treatment services within their communities, which are more affordable than those offered by tertiary hospitals, thereby lowering the financial burden of healthcare. Second, PHC services can help reduce the risk of disease, detect diseases at early stages, and get accessible health care services. All of these factors contribute to better health outcomes.

However, due to limitations of resources, salaries, and career development prospects, most medical students are less likely to pursue careers as general practitioners or family doctors in PHC institutions. ⁶ As a result, PHC institutions in many countries face a shortage of qualified healthcare professionals, which hinders the improvement of diagnostic and treatment capability and ultimately affects the overall effectiveness of PHC. ⁷ Nowadays, AI has rapidly permeated various aspects of daily life, offering a promising solution to these challenges. AI-based medical products can assist general practitioners and family doctors in analyzing patients’ health indicators, chief complaints, and examination results. ⁸ They provide suggestions for screening, diagnosis, treatment, and referral, thereby enhancing the efficiency and effectiveness of PHC. ⁹ Consequently, an increasing number of researchers view AI as a transformative force in the domain of PHC, and have conducted studies on its application in PHC.

Bibliometric analysis is a method that quantitatively assesses the existing body of literature on a specific topic using mathematical and statistical methods. ⁹ This method can visually present publication times, countries, institutions, authors, journals, references, and keywords in the extant studies, providing a comprehensive overview of trends, frontiers, and hotspots of the focal topic.^10,11

The value of bibliometric analysis has been demonstrated across various fields, including medicine and artificial intelligence.^11,12 Existing bibliometric studies have synthesized research on AI in healthcare. However, most of these studies focus on AI in the overall healthcare domain,^13–16 while few have specifically examined AI in PHC. PHC is represents a distinct sub-domain within healthcare, serving as the initial point and most accessible healthcare service for the majority of patients. It plays a critical role in the healthcare system due to its high service demand and its function as the first contact point for medical care. Despite its importance, PHC institutions often face a shortage of highly qualified physicians, as medical graduates are less inclined to pursue careers as general practitioners or family doctors. AI offers promising solutions to alleviate this shortage by improving efficiency and decision-making in PHC settings. Therefore, investigating AI in PHC is both meaningful and valuable. Although numerous studies have discussed AI use in healthcare, bibliometric analyses focusing specifically on AI in PHC remain scarce.^13–16 This bibliometric study aims to fill this gap by systematically mapping the existing literature on AI in PHC to identify its research trends, research distributions, hotspots, and emerging frontiers.

Method

Data extraction

This is a bibliometric study was conducted from February 2025 to June 2025. Tools including CiteSpace, VOSviewer, and Bibliometrix were employed to analyze papers sourced from the Web of Science Core Collection (WOSCC) database.

According to the BIBLIO checklist, ¹⁷ the identification and extraction of data were carried out in four steps.

First, a comprehensive search was conducted in the WOSCC to obtain data for the bibliometric analysis. The WOSCC was chosen as the data source for three reasons. First, the WOSCC is a comprehensive and standardized database recognized for its broad coverage across diverse research domains. ¹⁸ Second, bibliometric studies require consistency in the format of exported records, whereas records exported from different databases have different formats, making them unsuitable for integration. Third, the WOSCC is widely adopted in previous bibliometric studies.^10,11 The search query used in the WOSCC was TS = (“artificial intelligence” OR “machine intelligence” OR “machine learning” OR “deep learning” OR “supervised learning” OR “unsupervised learning” OR “semi-supervised learning” OR “reinforcement learning”) AND TS = (“primary healthcare” OR “primary care” OR “general practice” OR “general practitioner” OR “family medicine” OR “family doctor” OR “family physician”). The search period was set from January 1, 2015, to December 31, 2024. We limited the type of literature to ARTICLE and REVIEW, and restricted the language to English. After removing duplicated papers, 1308 papers were obtained.

Second, two researchers independently screened the titles and abstracts of these papers according to following eligibility criteria: (1) the paper was published between 2015 and 2024; (2) the paper was written in English; (3) the type of the paper was article or review; and (4) The paper was related to AI in PHC. In this step, 576 papers were excluded, resulting in 732 papers for further assessment.

Third, the full texts of the remaining papers were downloaded, and two researchers independently assessed the full texts according to above eligibility criteria. In this step, 79 papers were excluded, resulting in 653 papers included in the final bibliometric study.

The process of paper searching and selection is shown in Figure 1.

Figure 1.

Process of searching and selecting papers.

Fourth, for each of the 653 included papers, the following information was extracted: publication year, author, country, institution, journal, citation count, keywords.

Data analysis

CiteSpace, developed by Chaomei Chen, is a widely used software tool for bibliometric analysis. ¹⁹ In this study, CiteSpace 6.3.R1 was used to visualize the dual-map overlay of journals, cluster keywords, and literature bursts. VOSviewer, developed by Nees Jan van Eck et al., is another commonly used tool for generating bibliometric network visualizations. ²⁰ This study employed VOSviewer 1.6.20 to visually analyze and present the distribution, collaboration, and co-occurrence networks of countries, institutions, authors, journals, and keywords. In addition, Bibliometrix was utilized as a supplementary tool to visualize the distribution and collaboration of countries. ²¹ Finally, Origin and Pajek 5.19 were applied to further refine the visualizations of country collaboration networks and keywords co-occurrence.

Results

This study included 653 papers published between 2015 and 2024. A total of 4304 authors from 1551 institutions in 70 countries contributed to these papers.

Time trend of publications and citations

Time trend of publications and citations reflects the academic attention paid to AI in PHC over different periods. Figure 2 shows the annual publication volume and citation frequency for the included papers from 2015 to 2024. In general, the number of annual publications on AI in PHC has shown an upward trend, with the exception of a slight decline in 2017. Prior to 2018, annual publication volumes were relatively low, with no more than 7 papers published each year. Since 2018, the annual publication volume started to grow rapidly, peaking in 2024, which accounts for nearly one-third of the included papers. Citation frequency has also increased consistently over the same period. Each year from 2015 to 2024 has seen a rise in the number of citations for papers on AI in PHC. The year with highest annual citation frequency was 2024, with 3416 citations of included papers. To better understand the growth pattern of research on AI in PHC, the annual growth rates of publication volume and citation frequency were also calculated and marked in Figure 2. The value outside the bracket represents the annual growth rate of publication volume, while the value inside the bracket indicates the annual growth rate of citation frequency. It can be observed that before 2018, both publication volume and citation frequency showed the highest annual growth rates. However, due to their low base numbers, the annual increases were relatively small. From 2019 to 2021, the annual growth rates of both publication volume and citation frequency slightly declined, but their annual increase rates rose due to higher base numbers. From 2022 to 2024, as the base numbers of publication volume and citation frequency reached higher levels, their annual growth rates further slowed, yet their annual increments remained at a high level.

Figure 2.

Annual publication volume and annual citation frequency.

Countries distribution

Publication volume by country helps identify productive and influential countries in the domain of AI in PHC, and the collaborative relationships between countries reflect the state of cross-national cooperation in this field. A total of 70 countries contributed to research on AI in PHC. Figure 3(a) shows the publication volume associated with each country, with links indicating patterns of international collaboration. Most of the contributing countries are located in the Northern Hemisphere and the Western Hemisphere. The international collaborations are primarily concentrated between North America and Europe, North America and Oceania, as well as North America and East Asia. Africa is under-represented in the included articles, which is evidenced by the fact that most African countries have not published English articles indexed in WOSCC. Collaboration gaps are evident in South America and Africa, which is characterized by limited collaborative research between these regions and productive regions, such as North America and Europe.

Figure 3.

Analysis of countries participating in research about AI in PHC.

Table 1 presents the top 10 countries in terms of publication volume and citation frequency. The USA (223) and the United Kingdom (128) each contributed to more than 100 papers on AI in PHC, followed by China (79), Canada (60), and Australia (45). The USA recorded the highest citation frequency (6628), followed by the United Kingdom (3255) and Austria (1247). Figure 3(b) illustrates the international collaborations among the top 15 countries based on total link strength. The strongest collaborative tie is observed between the USA and the United Kingdom, followed by the USA and China. It can be seen that the USA is the leading country in the domain of AI in PHC, with the highest publication volume, citation frequency, and collaborative strength.

Table 1.

Top 10 countries in terms of publication volume and citation frequency.

Rank	Country	Publication	Country	Citation
1	USA	223	USA	6628
2	United Kingdom	128	United Kingdom	3255
3	China	79	Austria	1247
4	Canada	60	Singapore	1160
5	Australia	45	India	1076
6	Germany	36	China	1055
7	Spain	36	Australia	731
8	Netherlands	35	Canada	657
9	Sweden	27	Germany	647
10	South Korea	24	Spain	516

Institutions distribution

Publication volume of institutions reflects productive and influential institutions in the domain of AI in PHC. The top 10 institutions ranked by publication volume and citation frequency are shown in Table 2. The University of Toronto contributes the highest number of publications (n = 20) on AI in PHC, followed by Stanford University, University of Oxford (16), Harvard Medical School (15), and Imperial College London (15). Among the top 10 institutions by publication volume, five are based in the United Kingdom, and three are from the USA. The institution with the highest citation frequency is University of California, San Francisco (1120), followed by University of Washington (980), National University of Singapore (940), Sankara Nethralaya (908), and Oregon Health and Science University (869). Notably, eight of the top 10 most-cited institutions are based in the, reflecting the country's significant academic influence in the field of AI in PHC.

Table 2.

Top 10 institutions in terms of publication volume and citation frequency.

Rank	Institution	Publication	Institution	Citation
1	University of Toronto	20	University of California, San Francisco	1120
2	Stanford University	17	University of Washington	980
3	University of Oxford	16	National University of Singapore	940
4	Harvard Medical School	15	Sankara Nethralaya	908
5	Imperial College London	15	Oregon Health and Science University	869
6	Kings College London	14	University Iowa	846
7	Karolinska Institute	13	University of North Carolina	844
8	Mayo Clinic	13	Boston Biostatistics Research Foundation Inc	839
9	University of Cambridge	12	Veterans Administration Medical Center	839
10	University College London	12	Emmes Corporation	838

Analyzing co-occurrence and co-citation of institutions helps identify leading institutions of AI in PHC and potential opportunities for collaboration. Figure 4(a) shows institutional collaboration networks. Employing VOSviewer, institutions that published at least 2 papers on AI in PHC can be divided into eight groups. Figure 4(b) describes the proportion of papers published by each institution relative to the total number of publications during the past five years from 2020 to 2024. A color gradient was applied: yellow indicates a higher proportion, suggesting these institutions has been more active in recent years, while purple indicates a lower proportion, reflecting less recent activity. It can be seen that University of Sydney, Karolinska Institute, and the University of California, San Diego have conducted increasing number of relevant research in recent years. In contrast, University of Washington, University of Oxford, University College London, and Leiden University have published relatively fewer studies on AI in PHC during the same period.

Figure 4.

Analysis of institutions participating in research about AI in PHC.

Journals distribution

Bradford's law is often applied to describe the distribution of papers across different journals, helping to identify the core journals in the domain of AI in PHC. ²² According to this law, journals publishing papers on a focal topic can be divided into three zones. Journals in each zone publish approximately one-third of the total publications. ²² The number of journals in these zones follows a ratio of 1:a:a². ²² Journals in the first zone are core journals, those in the second zone are relevant journals, and those in the third zone are peripheral journals. ²² Using the Bibliometrix tool, we identified the core journals in the field of AI in PHC, as listed in Figure 5.

Figure 5.

Core journals publishing papers about AI in PHC.

The publication volume of journals reflects which journals focus most on AI in PHC. Table 3 presents the Top 10 journals ranked by publication volume. Scientific Reports (24) has published the highest number of papers, followed by BMJ Open (19) and PLoS One (18). All 10 journals are indexed in the first or second quartile of the Journal Citation Reports (JCR), indicating that AI in PHC has received considerable academic attention from influential journals.

Table 3.

Top 10 journals in terms of publication volume.

Rank	Journal	Publication	JCR IF (2024)	JCR Quartile (2024)
1	Scientific Reports	24	3.9	Q1
2	BMJ Open	19	2.3	Q2
3	PLoS One	18	2.6	Q2
4	Journal of Medical Internet Research	13	6.0	Q1
5	BMC Medical Informatics and Decision Making	13	3.8	Q2
6	JMIR Medical Informatics	11	3.8	Q2
7	BMC Primary Care	10	2.6	Q1
8	Diagnostics	9	3.3	Q1
9	Journal of the American Board of Family Medicine	9	2.6	Q1
10	Npj Digital Medicine	8	15.1	Q1

The citation frequency of journals reflects which journals are the most influential in the domain of AI in PHC. Table 4 shows the Top 10 journals ranked by citation frequency. The journal with the highest citation frequency is Npj Digital Medicine (926), followed by British Journal of Ophthalmology (815), Nature Medicine (710), Lancet Digital Health (579), and Circulation (567), each receiving more than 500 citations. Nine of these 10 journals are indexed in the first quartile of JCR, and six have impact factors above 5, indicating the strong academic influence of these studies.

Table 4.

Top 10 journals in terms of citation frequency.

Rank	Journal	Citation	JCR IF (2024)	JCR Quartile (2024)
1	Npj Digital Medicine	926	15.1	Q1
2	British Journal of Ophthalmology	815	3.5	Q1
3	Nature Medicine	710	50.0	Q1
4	Lancet Digital Health	579	24.1	Q1
5	Circulation	567	38.6	Q1
6	Scientific Reports	417	3.9	Q1
7	Journal of Medical Internet Research	371	6.0	Q1
8	PLoS One	324	2.6	Q2
9	JAMA Network Open	280	9.7	Q1
10	Journal of the American Medical Informatics Association	205	4.6	Q1

The clustering analysis of journals reveals which fields’ journals are most likely to publish research on AI in PHC, helping researchers identify sources for relevant references and appropriate venues for manuscript submission. The clustering analysis of journal is shown in Figure 6(a). It can be seen that, these journals are divided into three groups. The red cluster consists of comprehensive journal, such as Scientific Reports and BMJ Open. The green cluster includes journals primarily focused on disease diagnosis, such as Diagnostics, Journal of Glaucoma, and British Journal of Ophthalmology. In the blue cluster, there are journals focusing on the application of digital technology in the medicine (Digital Health, Journal of Medical Internet Research, etc.) or primary health care (BMC Primary Health, Annual Family Medicine, etc.). In Figure 6(b), these journals are categorized into six clusters based on co-citation frequency. The red cluster includes journals concentrating on the digital technologies in medicine, such as Journal of Medical Internet Research, BMC Medical Informatics and Decision Making, and International Journal of Medical Informatics. The green cluster comprises journals emphasizing mental health, such as Neuroimage, Neurology, Alzheimers and Dementia, and Depression and Anxiety. The blue cluster contains highly influential general medical journals, such as Nature Medicine, JAMA, and New England Journal of Medicine. The yellow cluster features a number of journals on dermatosis, including Journal of the American Academy of Dermatology, Archives of Dermatology, and British Journal of Dermatology. The light green cluster includes journals on cancers, such as Cancers and Frontiers in Oncology. The purple cluster consists of journal concentrating on Respiratory, including Respiratory Medicine and European Respiratory Journal.

Figure 6.

Analysis of journals publishing papers about AI in PHC.

The dual-map overlay of journals illustrates which fields are most engaged with AI in PHC, as well as the disciplinary foundations underlying existing studies. ²³ The dual-map overlay of journals is shown in Figure 6(c). The labels on the left side of the map represent the citing journals, while the those on the right side denote the cited journals. ²³ It can be seen that the citing journals are mainly from the field of MEDICINE, MEDICAL, and CLINICAL, which represent the research frontiers. The cited journals are mainly from MOLECULAR, BIOLOGY, GENETICS, HEALTH, NURSING, MEDICINE, PSYCHOLOGY, EDUCATION, SOCIAL, which refers to the knowledge base.

Authors analysis

Analysis of authors helps identify influential researchers and shared research interests within a specific field. Lotka's law indicates that most scholars in a field have published only one related paper, while those who publish multiple papers represent a smaller proportion of the total. ²⁴ Specifically, the number of researchers who have published N related papers is approximately 1/N² of those who have published just one. ²⁴ Figure 7 shows the proportion of researchers by the number of papers they have published on AI in PHC. It can be seen that the actual distribution of researchers in terms of related publication volume (red line) approximately aligns with the pattern described by Lotka's law (origin range). The majority of researchers in this field have authored only one or two related papers, and only about 2% of researchers have published three or more related papers.

Figure 7.

Author distribution in terms of publication volume.

The publication volume and citation frequency of authors reflect which scholars are productive and influential in the field of AI in PHC. Table 5 presents the Top 10 authors based on publication volume and citation frequency. The authors with the highest number of publications are Bazemore Andrew (7) and Vidal-Alaball Josep (7), followed by Farooqui Usman (6), Hill Nathan (6), and Kueper Jacqueline (6). Six of the Top 10 most prolific authors are from the United Kingdom, and three are from the USA. The author with the highest publication count is Peng Lily (1109), followed by Abramoff Michael (920) and Raman Rajiv (908), each with more than 900 citations. Among the top 10 most cited authors, six are affiliated with institutions in the USA, and three are from the Singapore.

Table 5.

Top 10 authors in terms of publication volume and citation frequency.

Rank	Author	Publication	Country	Author	Citation	Country
1	Bazemore, Andrew	7	USA	Peng, Lily	1109	USA
2	Vidal-Alaball, Josep	7	Spain	Abramoff, Michael.	920	USA
3	Farooqui, Usman	6	UK	Raman, Rajiv	908	India
4	Hill, Nathan.	6	UK	Lee, Aaron.	884	USA
5	Kueper, Jacqueline.	6	Canada	Wong, Tien Yin	850	Singapore
6	Liu, Yun	5	USA	Tan, Gavin Siew Wei	844	Singapore
7	Gordon, Jason	5	UK	Ting, Daniel Shu Wei	844	Singapore
8	Pollock, Kevin.	5	UK	Folk, James.	839	USA
9	Sandler, Belinda	5	UK	Lavin, Philip.	839	USA
10	Lin, Steven	5	USA	Birch, Michele	838	USA

The co-occurrence network of authors reflects the collaboration relationships among scholars in the domain of AI in PHC. Figure 8(a) shows the co-occurrence network of authors. Co-occurrence refers to instances where two or more authors collaborate on the same publication, which helps to reveal collaboration patterns and identify leaning figures in the field. Authors can be divided into 11 groups based on their co-occurance. Tai-seala Ming and Liu Yun are the centers of the two largest groups respectively. There are no connections between the different groups. Figure 8(b) displays the co-citation network of authors. Co-citation means that the papers of two authors are simultaneously cited by other authors in the same publication, which reflects the degree of similarity in their research focus. Authors are grouped into five groups based on co-citation. There are relatively dense links among these groups, indicating that the research interests of authors in this field are highly convergent.

Figure 8.

Analysis of authors participating research about AI in PHC.

Reference analysis

Citation frequency of references reflects the theoretical foundations and seminal literature on AI in PHC, helping researchers identify essential papers for reading and citation. Table 6 presents the Top 10 most cited papers on AI in PHC. All of these papers were published in high-impact journals, each with more than 100 citations. The most cited paper is “Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices” written by Abramoff, Michael et al. and published in NPJ Digital Medicine in 2018, with 838 citations. This paper introduces an AI system for detecting diabetic retinopathy in PHC and demonstrates the its high accuracy in this setting. This is one of the earliest clinical applications of AI in PHC. The most recent paper among the top 10 is “Artificial intelligence and machine learning algorithms for early detection of skin cancer in community and primary care settings: a systematic review” written by Owain Jones et al. and published in Lancet Digital Health in 2022, with 122 citations. This paper provides a systematic review of studies examining the application of AI for early detection of skin cancer in the PHC context.

Table 6.

Top 10 papers in term of citations frequency.

Rank	Year	Title	Journal	Citation
1	2018	Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices	NPJ Digital Medicine	838
2	2019	Artificial intelligence and deep learning in ophthalmology	British Journal of Ophthalmology	791
3	2018	Fully Automated Echocardiogram Interpretation in Clinical Practice: Feasibility and Diagnostic Accuracy	CIRCULATION	567
4	2020	A deep learning system for differential diagnosis of skin diseases	Nature Medicine	400
5	2021	Artificial intelligence-enabled electrocardiograms for identification of patients with low ejection fraction: a pragmatic, randomized clinical trial	Nature Medicine	224
6	2019	Artificial Intelligence and the Future of Primary Care: Exploratory Qualitative Study of UK General Practitioners’ Views	Journal of Medical Internet Research	143
7	2019	Development of high-throughput ATR-FTIR technology for rapid triage of brain cancer	Nature Communication	140
8	2020	A deep learning algorithm to detect chronic kidney disease from retinal photographs in community-based populations	Lancet Digital Health	134
9	2021	Predicting the risk of developing diabetic retinopathy using deep learning	Lancet Digital Health	123
10	2022	Artificial intelligence and machine learning algorithms for early detection of skin cancer in community and primary care settings: a systematic review	Lancet Digital Health	122

Co-citation analysis was conducted to describe the knowledge structure of AI in PHC. Figure 9(a) and (b) show the results of co-citation analysis. Based on the associations among references, the literature can be grouped into 12 groups. The earlies groups are # 2 (atrial fibrillation), # 4 (melanoma), and # 12 (spirometry). The most recent groups are # 5 (dermatology), # 8 (chatgpt), and #9 (diagnostic tests). Groups # 1 (diabetic retinopathy) and # 7 (transfer learning) appear to have evolved from # 0 (health equity). There is no evolution relationship among other groups, indicating relatively low level of homogeneity across different research areas represented in the reference network.

Figure 9.

Analysis of references related to AI in PHC.

References with citation bursts are those whose citation counts experience a significant increase over a certain period, reflecting indicating which knowledge attracted the most attention in AI in PHC during different periods. The Top 25 references with the strongest citation bursts are shown in Figure 10. The reference with the strongest citation burst is “Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs” by Gulshan et al., published in JAMA in 2016, ²⁵ followed by “Dermatologist-level classification of skin cancer with deep neural networks” by Esteva et al., published in Nature in 2017. ²⁶ Both of these references experienced citation bursts during the period from 2018 to 2021. Notably, five references have maintained citation bursts continuing through to 2024, such as “The practical implementation of artificial intelligence technologies in medicine” by He et al., published in Nature Medicine in 2019, ²⁷ and “The Value of Automated Diabetic Retinopathy Screening with the EyeArt System: A Study of More Than 100,000 Consecutive Encounters from People with Diabetes” by Bhaskaranand et al., published in Diabetes Technology & Therapeutics in 2019. ²⁸

Figure 10.

Top 25 references in terms of citation burst.

Keywords analysis

Keyword frequency helps to initially identify the initial research hotspots in AI in PHC. The Top 20 keywords are shown in Table 7. The most frequent keyword is “diagnosis” (63), followed by health (51), risk (50), prevalence (47), validation (42), and primary-care (41).

Table 7.

Top 20 keywords in terms of frequency.

Rank	Keywords	Frequency	Rank	Keywords	Frequency
1	Diagnosis	63	11	Disease	25
2	Health	51	12	Prediction	23
3	Risk	50	13	Cancer	21
4	Prevalence	47	14	Risk-factors	21
5	Validation	42	15	Health-care	19
6	Primary-care	41	16	Performance	19
7	Classification	39	17	Outcomes	18
8	Management	32	18	Depression	17
9	Artificial-intelligence	29	19	Impact	17
10	Care	29	20	Population	17

The keyword clustering network reflects the main topics around which existing research on AI in PHC has been conducted. Further analyses of the network's density and temporal distribution of the clustering network can reveal the relative importance and evolution of these topics.

Figure 11(a) shows the clustering network of keywords. Based on their co-occurrence, the keywords can be grouped into five clusters. The red cluster is related to diagnosis with the help of machine learning-trained systems (machine learning, diagnosis, prediction, etc.). The blue cluster is about family medicine empowered by products helping family doctor to provide health management service for individuals (health, family medicine, chatbot, large language model, etc.). The yellow cluster is relevant to the analysis of medical images with deep learning-based systems (deep learning, convolutional neural-network, etc.). The green cluster is associated with the mental health service with AI (mental health, depression, Alzheimers-disease, etc.). The purple cluster is mainly about the early screening of disease with AI (risk, risk factors, risk stratification, etc.). Figure 11(b) shows the density of the clustering network. Brighter areas of the network indicate topics and keywords addressed by a larger number of publications. It can be observed that diagnosis is the most frequently studied topic in existing research, whereas early screening has received the least attention. The topics of family medicine, mental health, and medical image analysis have received a moderate level of attention. Figure 11(c) shows the temporal distribution of keywords in the network, reflecting the evolution of topics. In the figure, darker-colored keywords appeared earlier in existing studies, while lighter-colored ones appeared more recently. It can be observed that research on diagnosis emerged first, followed by mental health. These two topics primarily concern the diagnosis of physical and mental diseases. Family medicine and medical image analysis are the two most recently emerging topics. In the context of family medicine, AI enables family doctors to monitor residents’ health status and to more effectively guide them in developing healthy lifestyles. In the context of medical image analysis, AI assists PHC physicians in processing and interpreting complex information in medical images, thereby enhancing their ability to diagnose complex diseases.

Figure 11.

Clustering analysis of keywords.

Discussion

In this bibliometric research, CiteSpace 6.3.R1, VOSviewer 1.6.20, and Bibliometrix were employed to analyze and visualize 653 papers on artificial intelligence in primary health care, published between 2015 and 2024 in the Web of Science Core Collection, aiming to identify the trends, distributions, frontiers, and hotspots of the domain.

General information

The bibliometric research is based on 653 relevant papers published by 4304 authors affiliated with 1551 institutions, indexed in the WOSCC over a 10-year period from January 1, 2015, to December 31, 2024. The time trend of publication volume reveals a rapid increase, particularly after 2019. Before that year, the field remained in its nascent stage, lacking a solid research foundation. Since 2019, there is a surge in the volume of related publications, indicating that AI in PHC has gained growing academic interest in recent years.

According to the visualization analysis of countries and institutions, the USA and the United Kingdom are the most productive countries in this field. This is largely due to the concentration of prolific institutions in these two countries—three of the top 10 most prolific institutions are based in the USA and five in the United Kingdom. Collaborative research is common, with the USA serving as the central hub, followed by the UK, highlighting their dominant roles in shaping the field. The majority of these collaborations have taken place among countries in North America, Europe, and Oceania. Although the shortage of highly qualified physicians in PHC institutions is more pronounced in developing countries, increasing their need for AI, publications on AI in PHC from these regions, especially in Asia, Africa, and South America, remain limited. In addition, cross-national collaborative studies on this topic are mainly conducted among developed countries, whereas collaborations among developing countries or between developing and developed countries are notably insufficient.

According to the analysis of journals, the Scientific Reports published has published the most papers on AI in PHC, and is also in the top 10 journals in terms of citation frequency. Npj Digital Medicine ranks first in total citations of relevant papers and is likewise among the top 10 most prolific journals. Scientific Reports, PLoS One, Journal of Medical Internet Research, and Npj Digital Medicine appear in both the top 10 most prolific and the top 10 most cited journals, suggesting their central role and strong influence in the field of AI in PHC. Most of these popular journals are in the medical domain, which aligns with the results of the dual-map overlay analysis of journals.

In terms of author analysis, Bazemore Andrew from the USA and Vidal-Alaball Josep from Spain have published the highest number of papers on AI in PHC. Peng Lily's research on this topic has been mostly cited, indicating his outstanding influence in this field. It is worth noting that there is no overlap between the top 10 authors by publication volume and the top 10 by citation frequency. This indicates that productivity in this field does not necessarily equate to academic impact, emphasizing the need for researchers to focus on both scientific production and contributions in future studies.

An analysis of the included papers and their references reveals that the top 10 most cited studies on AI in PHC mainly focus on using AI to diagnose different kinds of diseases, such as diabetic retinopathy, skin diseases, brain cancer, chronic kidney disease, low ejection fraction, and ophthalmology. In the top 10 papers, there are also papers exploring the use of AI for medical image analysis or doctors’ attitudes toward AI. The earlies citation bursts occurred in four references in 2018, reflecting that academic interest in AI in PHC began to intensify around this time, which aligns with the time trend of publication. It is notable that the citation bursts of five references have continued through the end of 2024, indicating that AI in PHC remains a dynamic research frontier attended by scholars.

Frontiers and hotspots

According to the co-occurrence analysis, the keywords from the included papers can be categorized into five groups, each representing a major research frontier or hotspot in the field of AI in PHC. These thematic areas may serve as valuable references for future studies.

The first major focus is the use of AI to assist diagnosis.^29–31 Many countries, especially developing countries, are facing challenges related to the uneven distribution of healthcare professionals. The number of highly skilled doctors is limited, many prefer to work in tertiary hospitals rather than in PHC institutions. ³² Therefore, PHC institutions often lack adequate human resources to deliver appropriate and timely care.^33,34 In this context, AI-driven decision support systems can significantly enhance diagnostic capabilities in PHC by predicting diseases of patients and recommending treatment plans.^35–37 These systems can provide ranked disease lists as decision-making references for doctors to support clinical decisions, thus reducing diagnostic errors in PHC. ³⁸ Compared to doctors in tertiary hospitals, doctors in PHC institutions often have more limited diagnostic experience and clinical expertise. AI-driven decision support systems are trained through machine learning using a large number of real cases, so the accuracy of its diagnosis and the effectiveness of its treatment plans are promising and can compensate for the limitation of doctors’ knowledge and skill. ³⁶ Therefore, such empowerment from AI can help them improve the quality of their diagnostic and treatment services as well as their own confidence.^39,40 In this way, AI holds promise in strengthening the capacity of PHC systems and mitigating the long-standing issue of healthcare workforce imbalance.

Second, family medicine is another important application context of AI in PHC. ⁴¹ In many regions, a substantial portion of the population does not have an assigned family doctor to provide health management services. In addition, a family doctor often needs to serve a number of patients, limiting the effectiveness of the health management services they provide. Due to the limitation of time and effort, family doctors are unable to provide in-depth and personalized services for each registered resident. AI-driven products, such as wearable devices and mobile applications, can address this issue by supporting family doctors and residents.^41–43 Wearable devices can continuously monitor vital health indicators like heart rate and blood pressure in real time. Mobile applications can analyze the data recorded by wearable devices,^44,45 enabling the systems to assess users’ health conditions, and provide recommendations related to exercise, diet, sleep, medication, and medical visits.^38–40 Furthermore, if a resident is identified as a patient of a certain disease, these mobile applications can promptly notify his/her family doctors to provide timely intervention or treatment. ⁴⁵ If residents do not ensure their health conditions, they can actively ask questions to these mobile applications. These applications, powered by large language models, can provide answers to help residents better understand their health conditions and corresponding reactions. This not only improves residents’ capability of health management but also reduces the burden of family doctors.^46,47

Third, AI also used to assist early screening of diseases in PHC. Unlike tertiary hospitals, where the primary focus is on disease treatment, PHC emphasizes disease prevention and health promotion. ⁴⁸ Currently, several AI-powered systems are being utilized to assess individuals’ health risks by analyzing factors such as medical history, health indicators, and lifestyle habits. These systems can detect early warning signs in otherwise asymptomatic individuals, enabling proactive interventions prior to disease onset.^49,50 By offering timely interventions, such as lifestyle recommendations or preventive treatments, AI can play a pivotal role in mitigating disease risk, ultimately helping individuals make informed decisions about their health. In comparison to the high costs associated with treating diseases after they occur, preventive measures facilitated by AI-driven systems not only are more cost-effective but also can significantly reduce the physical and emotional suffering on individuals.^51,52 Therefore, integrating AI-driven health risk prediction tools into routine health check-ups and disease screening programs within PHC settings can substantially lower medical expenses while simultaneously enhancing health outcomes. By facilitating early detection and intervention, AI supports a shift toward a more proactive and preventive model of healthcare, fostering the long-term health for the population.

Fourth, mental health service including predicting mental diseases and offering psychological counseling is also a critical application of AI in PHC. Mental health is a crucial aspect of overall well-being and an important task of PHC. AI-driven solutions can play a pivotal role in identifying and addressing mental health disorders. ⁵³ For instance, AI-based mental health products, including mobile apps, utilize advanced algorithms to analyze various user inputs, such as facial expressions, vocal tone, body movements, and language patterns. ⁵⁴ These apps can accurately assess users’ emotional states, and facilitate the early detection of potential mental health disorders such as stress, anxiety, or depression. Beyond initial assessments, these AI-driven products can also provide psychological support to individuals diagnosed with mental health disorders. By integrating large language models, such as ChatGPT and DeepSeek, these applications can engage users in therapeutic conversations, thereby offering personalized psychological counseling and mental health education.^55,56 These interactions can help mitigate emotional distress and reduce the risks associated with severe conditions such as suicidal ideation, depression, and anxiety. The ability of AI-driven systems to offer assessment can enhance accessibility to mental healthcare, particularly in areas with limited resources or where stigma surrounding mental health may prevent individuals from seeking help. Furthermore, by providing timely interventions and counseling, AI can contribute to improved mental health outcomes and support a preventive, proactive approach to care in PHC. In addition, AI tools such as ChatGPT can also be used to promote mental health education among children. These AI products can enhance the access of children and their families to health services and present health information in a more understandable and interactive manner to them. ⁵⁷ AI products, such as ChatGPT, can also provide real-time feedbacks to children. These functions of AI can promote effective and personalized mental health education. ⁵⁷

Fifth, medical image analysis is an emerging application of AI in PHC. Driven by advances in computing power and the increasing availability of training images, AI systems have demonstrated continuous improvement in their ability to interpret medical images. ⁵⁸ Early screening and diagnostics are important task of primary healthcare service.^59–61 The early screening and diagnostics for certain diseases, such as dermatology, diabetic eye disease, ⁶⁰ hip fracture, ⁶¹ and melanoma, ⁶¹ malignant lesions, ³⁹ require in-depth analysis of medical images. However, most PHC institutions lack the equipment and technical expertise to conduct such analyses. Doctors in PHC often do not receive specialized training in image interpretation, which can lead to missed opportunities for early diagnosis and intervention. To address this challenge, some AI-driven systems have been developed to automatically collect and analyze medical images. ⁴⁶ These systems enable PHC institutions to collect medical images of residents, and automatically analyze these images to initially assess the probability of specific diseases. ⁵⁶ If the initial assessment shows that a resident has a high likelihood of a certain disease, his/her medical image will be transmitted to specialists at tertiary hospitals for further evaluation. In this way, AI can empower doctors in PHC institutions to enhance the accuracy and effectiveness of early screening and diagnosis of complex diseases. This study focuses on PHC, a sub-domain of healthcare, and finds that deep learning-based medical image analysis is a recently emerging topic in this field. In contrast, existing bibliometric studies on the whole healthcare system indicate that deep learning has been a major topic in healthcare. ⁶² This suggests that AI-driven medical image analysis is relatively mature in the broader healthcare system, but its application in PHC remains in its early stages. Future research on AI in PHC could pay more attention to this topic.

The five topics identified in this study differ from those reported in existing bibliometric studies. Previous studies mainly focused on clinical decision support systems, health information systems, neural networks, predictive modeling, big data analytics, personalized medicine, diabetes mellitus, and chronic kidney disease. The primary reason for this difference might be that existing bibliometric studies largely examine AI applications across the overall healthcare domain. As a result, the topics they identified tend to be general technologies, products, or diseases within healthcare. In contrast, this study focuses specifically on AI in PHC. Since PHC is a sub-domain of healthcare, the search query used in this study differed, the number of included papers was smaller, and their content was more focused, enabling the identification of more specific topics. For example, family medicine, mental health, and early screening, which were identified in this study, are vital contexts in PHC but are less emphasized in large general hospitals. Another potential reason lies in the difference in time span. Most existing bibliometric studies covered publications up to 2019 or 2021, before or during the COVID-19 pandemic. This study, however, spans from 2015 to 2024, including the most recent publications from the post-pandemic years 2023 and 2024. This extended time span may also explain why this study identified topics not reported in previous bibliometric studies.

Practical implications

The findings of this study have practical implications for policymakers, academic organizations, and PHC institutions. This section accordingly provides suggestions for these stakeholders.

First, policymakers should support the research, development, and application of AI products designed for PHC. This study found that AI can help alleviate the shortage of highly qualified physicians in PHC, demonstrating promising application prospects and market potential. Policymakers could promote the development of AI products for PHC by adopting supportive policies such as investment incentives, tax reductions, financial subsidies, and relaxed entry requirements for relevant companies and research institutions. In addition, they could collaborate with practitioners and scholars to establish standards for AI products used in PHC, thereby encouraging companies to enhance product quality. Furthermore, policies ensuring data security, user privacy, and algorithmic fairness could increase the willingness of PHC institutions and physicians to adopt AI products in clinical practice.

Second, academic organizations should support research on AI in PHC in developing countries. PHC institutions in developing countries face more severe shortages of highly qualified physicians than those in developed countries and therefore have greater needs for AI support. However, this study found that publications on AI in PHC from developing countries are limited, possibly due to insufficient funding and lower international visibility of researchers from these regions. Accordingly, international academic organizations and institutions in developing countries should allocate funding to support scholars from these countries to conduct research on AI in PHC. At the same time, academic institutions in developed countries should introduce incentives to encourage collaborations between scholars from developed and developing countries on this topic.

Third, PHC institutions should encourage their physicians to integrate AI in their clinical practices. This study found that AI can empower PHC physicians in various tasks, including diagnosis, family medicine, risk prediction, mental health, and medical image analysis. PHC institutions should assess their weaknesses across these tasks and invest in AI products that enhance physicians’ capabilities in the corresponding areas. Moreover, PHC institutions should provide training programs to improve physicians’ AI literacy, enabling them to use AI tools more effectively and efficiently in their clinical practice.

Fourth, it is necessary to establish an AI-empowered PHC service system. Existing studies mainly focus on applying AI in specific PHC contexts, but have not further integrated these contexts into a comprehensive PHC service system empowered by AI. Some PHC services involve multiple contexts. For example, chronic disease prevention and control services encompass early screening, diagnosis, and family medicine. AI products used in these contexts are often developed by different companies, and these products have not yet achieved data sharing, which limits the effectiveness of AI-assistance. How to formulate policies and strategies to unify AI-empowered PHC contexts warrants further investigation.

Fifth, enhancing PHC physicians’ trust in AI products and improving their interaction experience with AI could facilitate the adoption of AI in PHC. AI adoption in PHC depends not only on technical performance but also on trust and human-AI interaction. Privacy concerns, ethical considerations, cultural resistance, algorithm reliability, and professional autonomy are potential barriers to AI adoption in PHC.^63,64 PHC institutions and relevant companies can facilitate AI adoption by enhancing transparency of AI products, providing training for physician, and implementing safeguards against over-reliance. ⁶³

Limitations and future research

This bibliometric study has some limitations.

First, all included papers were obtained from WOSCC. Although WOSCC is a highly regarded and comprehensive database that covers most influential literature, some relevant publications may not be indexed. Future research could expand the search to include more comprehensive or field-specific databases to capture a broader range of studies.

Second, the depth of analysis is limited. As a bibliometric study, this method is often used to identify or visualize the general trend, distribution, frontiers, and hotspots of publications on a given topic. Future research could utilize systematic reviews or meta-analyses to conduct a more in-depth examination and integration of literature on AI in PHC, providing richer insights for researchers and practitioners in the domain.

Third, this study was limited to English-language articles. As bibliometric analysis is suitable for articles in the same language, this study only included English articles from WOSCC. Therefore, some PHC practices in non-English-speaking regions might not be included. In the country distribution analysis, publications and collaborative research among developing countries were relatively limited. In the thematic analysis, we found that articles related to early screening were also relatively few. Including non-English articles may potentially alter these geographical and thematic patterns. Future bibliometric studies could adopt multilingual searches to include articles in different languages, so as to validate and extend the findings of this study.

Conclusion

In this study, 653 papers published between 2015 and 2024 on the topic of AI in PHC were obtained from WOSCC, and bibliometric analysis was employed to identify the trends, distributions, frontiers, and hotspots of this field. The volume of related publication witnessed a rapid growth since 2019, indicating increasing academic attention to the topic in recent years. Analysis of countries and institutions reveals that the USA and the United Kingdom are two leading centers contributors in this domain. Scientific Reports, PLoS One, Journal of Medical Internet Research, and Npj Digital Medicine demonstrate both high publications and high citations in this field. The earliest citation bursts in this field occurred in 2018, aligning with the time trend of publication volume. There are also several ongoing citation bursts, indicating scholars’ continued interest in this topic. Clustering analysis of keywords identified five major application settings of AI in PHC: diagnosis, health management, risk prediction, mental health assessment and consultation, medial image analysis, which are frontiers and hotspots in this field.