Screening Strategy

In this analysis, all journal papers about AI in health care were included for screening. The papers for analysis were restricted to those that (1) were written in the English language, (2) focused on promoting health or health care, and (3) involved AI technologies. As AI technology is a leading-edge and rapid update research area, papers published in peer-reviewed journals, conference proceedings, and early access articles were included. Book chapters and books were excluded from this bibliometric analysis.

The coauthors received training of bibliometric analysis screening by watching the video of Müller's study screening guide [15]. The screening procedure was conducted based on the screening guide [15]. Two coauthors (YG and ZH) conducted a pilot screening with the same set of 50 papers. During the pilot screening process, papers from the preliminary searches were categorized into include, exclude, or unsure. Papers marked as unsure were screened by 3 of the authors (YG, ZH, and FY) and discussed until consensus was reached in team meetings. The practical inclusion and exclusion criteria were determined. After that, 2 coauthors (ZH and YG) each received a unique set of papers for title and abstract screening. Different from a systematic review, a bibliometric analysis only requires screening of the abstract and full text when it is necessary. First, 32 inapplicable and duplicate papers were removed. According to the screening criteria, 3730 papers were excluded either because they did not focus on promoting health care or did not involve AI technologies. Finally, 1473 papers were included for bibliometric analysis (Figure 1).

Open in a separate window

Figure 1

Flowchart detailing the paper collection and screening process.

Publication Patterns

Overall, 95.59% (1408/1473) of the included papers were published from 10 countries, shown in Table 1. About 45.42% (669/1473) of the included studied were from the United States. China was the next leading country (183/1473, 12.42%), followed by England (113/1473, 7.67%). The papers were published in 715 different journals. As demonstrated in Table 2, PLOS One published the most papers (57/1473, 3.87%), followed by Artificial Intelligence in Medicine (24/1473, 1.63%), and Expert Systems with Applications (1.63%, 24/1473). WoS subject categories were employed to indicate the research domains of included publications, as seen in Table 3. The research domains of computer science (252/1473, 17.11%), engineering (192/1473, 13.03%), and medical informatics (169/1473, 11.47%) are the top research domains of health care–related AI research.

Characteristics of Research Activities

Keywords are the core word extractions provided by researchers in the studies. Table 4 shows information about the frequency and centrality of keywords. The top 5 health problems are cancer, depression, Alzheimer disease, heart failure, and diabetes. The top 5 AI technologies are machine learning, artificial neural networks, deep learning neural networks, electronic health records, and support vector machines. The top 5 functions are case classification, diagnosis, prediction, risk estimate, and chronic condition management. The top 5 populations focused on in health care–related AI studies are children, adults, women, men, and elderly persons.

HistCite intelligently classified the research topic into 12 clusters, labeled from 0 to 11 in Figure 3. The modularity (Q) was 0.423, which was higher than 0.3, indicating that the cluster results were significant. Cluster 0 is the largest cluster (coronary artery disease) and cluster 11 is the smallest one (diabetes mellitus). Each cluster was generated based on the number of keywords under one research domain, not the frequency of keywords.

Open in a separate window

Figure 3

Cluster analysis of artificial intelligence health care publications.

Growth Rate of Publications

Since the first publication in 1995, health care–related AI research had a slow increase in the following 17 years. However, since 2012, the field has started to move forward at a fast pace. In the past 5 years, the growth rate of publication reached 45.1%, which is more than 4 times the publication growth rate from 1995 to 2014. The recent rapid growth is due to multiple reasons. The technological breakthroughs of AI in this time period contributed to the explosive growth of AI adoption in health care–related AI research [17]. From 2015 to 2018, the appearance of autonomous robots, voice recognition, neural networks, and machine learning provided unprecedented opportunities for the prediction, diagnosis, and treatment or management of diseases [18,19]. According to the exponential growth pattern of health care–related AI research, the publications in health care–related AI will continue to grow in the future. Applying the growth rate of the past 5 years to the next 5 years, we can estimate that the publication volume in health care–related AI research will double about every two years.

Publication Patterns

The field of AI health care has attracted people from all around the world, but the high-income countries are the main force in health care–related AI research. The United States itself contributed about half of the research in the field of health care–related AI research. Compared with the rapid advancement of health care–related AI research, this research field in non–high-income countries is still meager. This observation of the small outputs in middle-income and low-income countries causes concern because many low-income countries have limited health care resources, but their public health issues are nonetheless increasing dramatically due to rapid globalization and urbanization [20]. Some developed countries have launched federal AI health care policies incorporating guidance on the development and regulation of AI in health care, such as the UK Code of Conduct for Data-Driven Health and Care Technology [21]. Consequently, about 85% of the research outputs were generated in developed countries, but 80% of the world population lives in developing countries. Multiple barriers, including funding, prioritization, research capacity, infrastructure, and language contribute to these disproportionate results [22]. From a global perspective, AI technologies are promising in terms of promoting health outcomes in low-income countries with limited health care resources.

Based on the output and citation counts, health care–related AI research is generally favored by large-scale journals related to health care. However, the sizes of the research effect are field-specific. The improvement and sophistication of the fields of computer science and engineering have paved the way for the development of AI. It is worth noting that the field of oncology received the second most citations. These results are consistent with the findings of the top keywords of included papers, implying that published AI papers on cancer care accumulated a high number of citations.

Characteristics of Research Activities

AI is redefining and disrupting the way health care is being carried out across diverse levels, based on the results of the characteristics of AI publications. According to the top keywords in the identified categories, the top domains of disease in AI research are cancer, coronary artery disease, chronic kidney disease, prostate cancer, and diabetes mellitus. The diseases of the leading causes of death received the major attention of AI research. By 2030, chronic diseases will contribute 80% of the human deaths globally and result in severe global burden of disease [23-25]. Specifically, the cancer mortality rate has steadily risen by 6% during the past 10 years [26]. Therefore, researchers direct their hopes and efforts on early detection and condition management by using advanced technology [27,28]. AI is making its way into cancer treatment from diagnostic classification to tumor behavior prediction [29,30]. Furthermore, cardiovascular diseases are also highly related to behaviors, especially among older adults [31,32]. AI can exert a proactive role in predicting risk factors for cardiovascular diseases related to older adults’ behaviors in order to remarkably reduce the hospitalization rate, readmission rate, cost of care, and even mortality rate [33-35].

Extensive functions of AI have assisted modern health care by providing smart medical data analysis and developing accurate and efficient prediction for treatment [36]. Machine learning is the most commonly used AI technique in health care. Artificial neural networks are the next biggest research area related to health care. There are promising applications of these AI technologies to the development of interventions for other diseases in future research, such as mental health interventions, health education, and chronic disease management.

AI research is witnessing widespread adoption in the prediction, detection, diagnosis, classification, treatment, and survival prediction of diseases [30,37]. The most common application of AI technologies is reflected in the domains of medical classification and quality of care. AI is investing significantly in improving the quality of care in the health care system [38]. The potential of AI in reviewing medical images and analyzing large-scale data has led to significant improvement in the quality of care [39]. Additionally, AI-based risk prediction models can investigate the complex relationships between clinical data and disease treatment [40].

Regarding the research populations, researchers seem more interested in child populations. Due to the development of cognitive aids to support diagnosis, treatment, care coordination, surveillance and prevention, and health maintenance, improvement of AI in clinical pediatric health care is remarkable [41]. However, cancer and various chronic diseases are the main focus of the current health care–related AI publications. Considering that elderly populations are the primary sufferers of chronic diseases, future studies may shift their angle to further explore AI implications on health in older adult populations.

Research Hotspot Tendencies

In the last three decades, the times and lengths of citation bursts of each health care–related AI research topic have varied. Certain keywords were extraordinary consistent for a long period of time, while some keywords only briefly surged in the field of AI health care. The period of 1997 to 2014 contained a concentrated outbreak of multiple health care–related AI study bursts (Figure 4). AI technology research involved more health care fields and produced more diverse keywords in recent years. It is possible that the more diverse keywords of health care–related AI research diluted the citation bursts.

According to the citation rates, artificial neural networks, support vector machines, and convolutional neural networks have the highest impact on health care. These techniques have been widely adopted by oncology research to predict disease and survival rates [42]. Historically, logistic analysis has been a widely used research method of health care–related AI. However, survival analysis has had a recent citation burst in this field.

The studies in these clusters focus on testing AI techniques and translating these techniques into practical settings [37,43]. The next step of health care–related AI research may transform from lab-based research to the development of clinically validated and safe regulated systems. The current health care–related AI research faces the challenges of deploying lab-based AI intervention into clinical practice [44]. These challenges include ethical issues brought up by AI application [45] and the quality and quantity of medical data and cases [46], which will decide the diagnosis and treatment ability of AI.

Health care–related AI research experienced several paradigm shifts during the past decades and continues to shift. As shown in the timeline overview (Multimedia Appendix 1), the sustainability of health care–related AI research clusters is influenced by the increasing availability of health care data and the rapid progression of AI techniques. The research of cancer, convolutional neural networks, and nucleosides sustained a hotspot over the past twenty years, whereas some clusters were relatively short-lived, such as chronic kidney disease, prostate cancer, and diabetes mellitus. Particularly, nucleosides, convolutional neural networks, and tumor markers have remained research hotspots through 2019. These research domains exert strong impacts on the field and these analysis results suggest that their influence will likely continue in the next few years.

Limitations

Limitations of our work need to be acknowledged. First, although we are confident that a single database—WoS—is a large enough database to offer a wide variety of publications vital for our analyses, future studies will apply other databases, such as Scopus, to explore more potential papers. Second, the search keywords (eg, mental health, behavioral health, and health care) related to health care were quite general, which may not be able to identify AI-related studies in all aspects of health care. For example, AI-related studies of HIV might have been omitted. Third, some keywords, although they were ranked as top keywords, were uninformative by themselves (eg, risk, model, and system) and could not be analyzed. Fourth, we did not include gray literature (eg, books), and we did not include papers published in languages other than English. Therefore, we may have missed relevant studies conducted in different forms, languages, and countries. Future studies can broaden the search scope to explore more relevant research to enrich the literature.