As seen in Table 1, the number of AI publications increased rapidly during the past 40 years. Notably, most of the publications (23,216 papers, 84.6%) were published during the last 10 years, and 60.6% of the total citation belonged to this period. The usage of papers was counted by the number of downloads. The mean use rate (download rate) within the last 6 months, of papers published in the year 2018 was three times higher than that of papers published in the previous years. The mean use rate within the last 5 years reached its peak in 2013 and decreased from 2012.

We analyzed the frequency of a country where the study was conducted, which was mentioned in the abstract (Table 2). Among 50 countries, the United States appeared the most (1867 times, 40.4%). Notably, only four African countries (Egypt, Niger, Kenya, and Nigeria) were mentioned in the abstracts. In addition, 13 Asian countries contributed to this list, and two Asian leaders of AI technologies—China (including Taiwan and Hong Kong) and India—accounted for 9.8% and 4.32% of the total papers, respectively.

Table 3 presents the scientific research topics constructed by LDA. By analyzing the most frequent words and titles, we could manually annotate the label of each topic. Robotics, which most mentioned the 10 topics and branches of AI (topic 1, topic 6, and topic 9), has supported surgery and treatment. AI types were applied the most in the diagnosis and prediction (topic 2, topic 5, and topic 7). Based on development visualization, there was a growing trend in some of the 10 topics, with different rates. The number of papers related to topic 1 was highest and increased gradually but with a slower rate in recent years. Moreover, the number of papers in topic 2 and topic 3 increased at a higher rate than that of papers in other topics (Figure 1).

Based on the classification of research areas in the Web of Science, we identified the dendrograms for the areas (Figure 2). The dendrogram includes the clades and leaves. The clade is the branch, and each clade includes one or more research areas. The horizontal axis shows the distance or dissimilarity between research areas. Each joining (fusion) of two clusters is represented on the diagram by the splitting of a vertical line into two vertical lines. The vertical position of the split, shown by a short bar, gives the distance (dissimilarity) between the two research areas. It shows that the AI applications focused on seven following research areas: surgery, robotics, and noncommunicable diseases (hepatocardiac disorders or cancer); neurosciences and psychiatry; the application of electronic health (telecommunication); chemical sciences; nanoscience; electrochemistry; and medical informatics and biotechnology. It seems that AI in medicine was assigned mainly to the disciplines diseases and treatment (surgery or robotics application).

We applied PCA to identify the landscape of AI in medicine and health care (Figure 3). Based on the size of the node, most papers belonged to the following research categories: clinical: surgery, radiology, and nuclear medicine; technology: biomedical, robotics, computer science, medical informatics; and diseases: oncology, general and internal medicine and noncommunicable diseases. As shown in Figure 2, a strong relationship among the applications of AI in treatment, diseases, and medical informatics shows that AI assisted surgeons, especially in some diseases for which surgery is key in treatment or diagnosis, such as cancer or cardiovascular diseases. The combination of information science, computer science, and health care, called health informatics, has created a wide range of applications, from cell level to population level [41]. Collision of several computer science–related fields and medical fields created a multidisciplinary science, which has led to better chances of providing the best treatment to patients. Additionally, the development of computer sciences has contributed to the advancement of AI in pharmacy, biotechnology, and chemistry in areas such as drug discovery, drug identification and validation, and drug trials.

We compared ten research topics by LDA (Table 4) with Web of Science research areas (Multimedia Appendices 3-5) to identify the consistency of research disciplinaries of AI in medicine and health care. Computer science and its related fields appeared the most (eight topics). The major application of computer science has been in medical fields: from cells (gene microbiology information, topic 5), disease (oncology, cardiovascular, topic 7), and diagnosis and treatment (topic 1, topic 6, and topic 9) to health policy (topic 3). Additionally, AI types were used the most in medicine and health care, including expert systems, artificial neural networks, machine learning, and natural language processing. Robot and surgery were two applications mentioned the most in topic 1, topic 6, and topic 9. Robotic-assisted procedures were used for cancer surgery and cardiovascular diseases. Robot-assisted therapy was used in treating sports concussion or neurorehabilitation (topic 9).

By analyzing scientific research publications and modeling their research topics, we generally described the 42-year development and identified the trend of AI application in medicine and health care. The mean use rate related to the application of AI in medicine was the highest in the last 5 years and tended to reduce since 2012. This can be explained by the rapid development of technology and research [42]: Scientific papers published more than 5 years ago would not attract the attention of scientists. Therefore, dissemination efforts need to be taken into consideration by not only policy makers but also authors, to increase the influence and implement changes in practice settings [43]. In addition, the results show a rapid increase in research productivity and downloaded papers in the last 5 years. Its growth was contributed mostly by western countries, driven by the United States. Among 11 Asian countries in that list, China and India were two leaders in research on AI in medicine. The application of AI has benefitted the health care system in high-income countries. One study showed that the United States could save US $5-$8 billion per year with the application of information technology in health care [44]. Another recent analysis found that with the application of AI, we can save up to US $150 billion in yearly health costs [45]. AI, however, has not been widely used in low-income countries. This could be due to the undeveloped infrastructure in the internet, technology, and health systems and a lack of highly qualified human resource. Regardless of the disadvantages, AI holds promise for changing health care services in low-income countries [46].

Based on the topics and research areas, we found that the application of AI in medicine and health care has been focused on robot support in surgery (topic 1) and rehabilitation (topic 9), AI in diagnosis and clinical decisions support (topic 2, topic 4, topic 5, topic 7, topic 8, and topic 10), and AI in health care system management (topic 3). First, for clinical treatment, our results confirm that medical robots and robot-assisted surgeries have been widely used [47,48]. AI has been widely applied in surgery due to its benefits for patients and medical professionals, such as increased accuracy, reduced operation time, minimized surgical trauma, and reduced length of recovery time for patients [49]. Second, AI methods such as machine learning and natural language process analyze complex medical data [20], decrease time spent finding relevant evidence, and reduce medical errors that improve the quality of diagnosis in medical health care [50]. Finally, AI will certainly be applied more in the health care system in the future owing to its advantages over the traditional decision-making process. On the other hand, the fact that users do not know how the results are analyzed by the “black box” algorithms, ethnic differences in validity of facial recognition technology for genetic diagnosis, medical and behavioral conditions [51,52], and ethnic bias in training data set [53] raise questions about product liability, privacy and data protection, and ethical and legal issues [51]. Thus, researchers have voiced their concern about legacy and ethical guidelines that are lagging behind the development of AI in health care and medicine [51].

Our findings have some implications for health research and policy. The quick development of AI applications in health and medicine requires some preparations. AI may change the relationship between caregivers and patients, as the direct interaction might reduce due to digital tools such as a free app in the patient’s personal device, which could diagnose the disease in some cases or even lead to self-diagnosis via the Web [54]. Thus, it is necessary for all parties involved to ensure that, in the case of mental health diagnosis, for instance, subtle signs of mental illness would not be neglected [55]. In addition, standard guidelines or laws about collecting private information or application of AI in all health care sectors are urgently needed [56], as the application of AI in health care and medicine has potential threats to patients’ privacy and safety. Finally, AI is transforming health care in resource-poor settings and reducing the gap between rural and urban areas [46]. In rural areas of developing countries, the shortage of medical doctors and trained nurses and the limitation of medical techniques and machines have reduced the quality of medical services [57]. In addition, it is difficult to attract skilled medical workers in rural areas due to the poor working environment and living conditions [58]. However, the development of AI applications can be a solution to these problems. For instance, the AI method (machine learning) proposed a model helping forecast dengue outbreaks in China [59]. In addition, AI has proven to be effective, with a high accuracy of breast cancer detection [60,61]. Moreover, AI can reduce medical costs in developing countries. For example, a highly effective AI method could provide an alternative to expensive diagnostic methods to classify acute leukemia [62]. However, absorptive capacity, local culture, legacy [63], and infrastructure (eg, electricity, internet, or financial source) should be carefully taken into consideration [64]. Notably, policy development for AI should be given more attention, since its failure has been recognized in developing countries such as Vietnam [65].

Our study has several limitations worth noting. First, we choose only Web of Science as the database, which may not cover all the publications in the research fields. Second, only English articles and reviews were analyzed in this study. Finally, we applied LDA to model the topic research in title and abstracts, not the full text. However, two other methods (coincidence analysis and PCA) confirmed similar results about the connections of research topics. Thus, LDA could be considered a support method to reduce the workload in the screening step for future systematic reviews [66].

The application of AI in medicine has grown rapidly and focuses on three leading platforms: clinical practices, clinical material, and policies. AI might be one of the methods to reduce the inequality in health care and medicine between developing and developed countries. Technology transfer and support from developed countries, along with the internal efforts of poor-setting countries, help in the development of AI applications in health care and medicine.