Research on artificial intelligence, machine and deep learning in medicine: global characteristics, readiness, and equity | Globalization and Health

Several bibliometric analyses have already been carried out, focusing on AI in medicine. In most cases, however, these relate to a specific medical topic [36,37,38] but not to the field of AI in medicine and healthcare as a whole. One study was conducted on AI_med. It only extended to 2018 and provided a first look at keywords, authors, and countries of origin [39]. The novelty of this study is, therefore, that it paints a comprehensive picture of global research on AI_med, interpreting socio-economic characteristics and, thus, highlighting particular country-specific research incentives and constraints.

Chronological aspects

The chronological analysis of global publication development shows connections with the history and development of AI_med. After a phase of initial enthusiasm about the possibilities of AI, disillusionment followed in the 1970s. In reference to the term “nuclear winter” from the Cold War era, the “AI winter” refers to the period in which AI activity in business and academia declined dramatically after the US government decided to scale back its involvement in AI research [40]. The previously overestimated promises of AI have not been fulfilled during this time. The 1973 Lighthill Report [41], initiated by the British Research Council, did its part by stating that no area of AI had had the major impact promised. Research output held correspondingly low during this period.

Nevertheless, AI did not find its way into medicine or the health sciences until the 1970s [42]. The first medical approaches published in WoS were developed to facilitate antimicrobial therapy for physicians [43] and pattern recognition in neurobiology [44]. In the 1980s and 1990s, little research was conducted that focused on faster data collection, surgical intervention support, database administration (DBA), and electronic health record (EHR) implementation [42].

The low level of annual numbers changed with the beginning of the 2000s, when research related to AI_med first increased slowly, followed by a sharp increase in 2017. Advances in the function of AI in medicine led to this development. The most cited articles on AI_med were published during this time, showing a strong interest in AI_med [45,46,47]. They are related to diagnosis with AI in medical imaging algorithms, gene expression profiling, and DL with convolutional neural networks. Before this sharp increase, another peak in the annual citation rate occurs. The year 2009 presented no highly cited articles but some well-recognized studies on pattern recognition of ML in MRI [48, 49]. However, the annual citation numbers started already to rise rapidly before the increase in publication numbers. Until now, they reached their highest annual value in 2020, when the most frequently cited articles were published. Since publications in the biomedical sciences usually take approximately eight years to achieve half of their citation (cited half-life: CHL) [50], this development underlines the extremely strong interest in the AI_med articles with a certainly increasing future trend.

However, the development of AI has been associated with high risks and problems, particularly in the field of medicine. In recent years, AI has also been used to create fake news on medical topics, creating and spreading confusion. This happened, for example, in the manipulation of people, which led to a worldwide movement against vaccinations [51]. The EU AI Act, which was introduced in 2024, was the first instrument to legally regulate AI and thus respond to the growing and irrefutable concerns and fears, including on the professional side [2].

Aspects of research foci

Similarly, Radiology/Nuclear Medicine is the most frequently addressed research focus for title and keyword analysis, demonstrating the importance of AI for the application of medical image patterns. Not surprisingly, the WoS categories Oncology, Cardiology, and Neurology are also very commonly addressed in AI_med. The prediction of mortality and morbidity is an important topic for genetics studies and is also underscored by frequently occurring terms. Cluster analyses revealed the deep interconnectedness of research patterns around the concepts of AI, ML, and DL.

The research foci of the publishing countries are mostly similarly distributed.

Comparing the two leading countries, AI research in China– in contrast to that in the USA– is more dominated by mathematical and computational aspects. China is also slightly stronger in articles on oncology, while its share of articles on medical imaging is lower. These two countries are by far the most active in the field of AI_med.

Geographical aspects

In line with the former studies [36,37,38], the USA and China published the most on AI_med, followed by some European countries. While Italy ranks among the top five countries in these previous studies, it ranks only 9th in our analysis. Italy’s share has decreased since 2016 when other European countries were doing more research on AI_med.

Until the 2000s, the USA was the absolute leader. China has been gradually catching up since then and has overtaken the USA since the 2020s. In terms of the number of citations, China is still far behind, with approximately 37% of the citations for US articles at the time of the evaluation. This is certainly due in small part to the higher proportion of mathematical and computational foci in Chinese articles, as these disciplines are generally cited comparatively less when looking at the Journal Citation Report (JCR) of the journals in question. However, this cannot be the sole reason for the lower citation numbers. China is also not involved in the most cited articles and ranks only 18th with its most cited study as a first-author country with a 2020 study on AI for COVID-19 detection [52]. This lower citation level is also underscored by the comparison between US and Chinese institutions. China has certainly gained a lot with its 2017 New Generation AI Development Plan [53]. This plan focuses primarily on promoting AI development but also initiates a high-level time frame for regulations, e.g., the Cybersecurity Law of 2017 [54]. Its introduction has led to a large increase in the number of related scientific studies in China. However, China also needs to pay attention to the quality of its projects. Only future analysis will show whether this plan has been or will be successful, both in terms of global recognition of scientific studies as measured by the number of citations. Regulatory principles, such as the guidelines for the registration and review of AI-based medical devices published by the National Medical Products Agency (NMPA) in 2022 [55], must prove their worth. It is said that policymakers in other countries can learn from these plans as they give a sneak of what might be possible and what might be futile. It must be taken into account that this development has its roots in China’s regulation of internet content, which illustrates the CCP’s (China Communist Party) political interest. Therefore, few actors will swim against the ideological tide, and opposing solutions will not be considered [53]. In comparison, there is currently no specific regulatory pathway for AI technologies in the USA. Here, the Food and Drug Administration (FDA) evaluates them under the framework conditions for other medical devices [56]. In addition, the EU and Brazil are attempting to regulate the development of AI on a risk-based basis, while most other countries are seeking national guidelines for AI applications in healthcare [55].

Economic and equity aspects

As we have shown, the level of economic power, innovation, and readiness for AI is significantly correlated with publication performance on AI_med worldwide. Previous studies on a variety of scientific topics have also shown that the economic strength of countries is linked to their publication performance [57,58,59]. Therefore, developing countries or countries with lower economic power are underrepresented in AI_med research. However, some of these countries have been able to perform well in terms of publication numbers, albeit at relatively low scores. These countries include Pakistan, Egypt, Bangladesh, and Ethiopia. Despite their low economic power, these countries can be singled out as examples of having good opportunities for AI development in the health sector. Fundamentally, countries in the Global South are in a worse position to benefit from AI technologies because they do not have the opportunity for equal progress and implementation. This may exacerbate global inequalities in the health sector as a whole [35]. In addition to Cyprus, which is in the lead with respect to the R_GDP analysis. Also, Jordan and Iran are among the top three ranks. Both countries are categorized as LMI economies [60].

Differences concerning the country of origin can also be shown for the research foci. India, for example, is the only country in the top ten that is not an HI economy. It has the lowest proportion of articles on medical imaging, but a higher proportion of articles on math, computing and engineering topics compared to the other top ten countries. These focus areas also refer to Jordan, Iran, and Pakistan’s foremost themes, engineering and mathematics. In contrast to Jordan and Pakistan, which mostly collaborate with Saudi Arabia, Iran is more scientifically linked to Western countries (the USA and Canada). India collaborates most frequently with the USA, but this is immediately followed by a collaboration with Saudi Arabia, which shows that this is an important scientific partner in AI_med research in these parts of the world. With Saudi Arabia’s comparable low citation rate, collaborations with Western HI countries lead to a higher reputation in the scientific community.

The differences between countries in terms of performance and readiness in AI are due to structural characteristics in infrastructure and research [61]. This is driven by different levels of development and funding, particularly in AI_med, which is additionally dependent on well-curated data [18, 39].

Affiliation aspects

This difference in the global recognition of research output on AI_med is also reflected in the metrics of the publishing institutions, which also show an inferiority of the Global South with significantly lower citation figures compared to Western countries. Despite its participation in highly cited papers [62, 63], the Canadian University of Toronto achieves values well below the average of the ten leading institutions. It is, therefore, an exception in terms of the citation frequency of AI_med publications. This may be due to the late increase in the use of its research in AI_med (2019), thus, better citation values can be expected in the future.

In contrast to general AI research, private companies are less involved in AI_med [64]. The US company Google LLC ranks first in general AI, even when looking at the joint ranking of science and industry. Google lost its pole place in AI_med, where it is ranked behind Siemens Corp (Germany), GE Healthcare (USA), and Philips Corp (UK). In the area of general AI research, Microsoft is in second place, followed by Facebook. GE Healthcare, which is frequently active in AI_med, is not included in the general AI list.

A comparison of the countries’ research performance on general AI and AI_med shows– similar to industrial research– that Germany is more active in the medical context. Instead, Israel is more involved in general AI, where it ranks third [65].

Future aspects

Countries around the world are currently striving to equip their scientific, bureaucratic, and industrial systems with AI capabilities, not only to exploit their potential but also to take a leading position in further development. Therefore, the next few years will show who will set the standards. However, the steps have been mapped out, and the risks cannot be denied, thus, most countries have already developed and will continue to develop strategies to tackle the problem [66].

The WHO has already emphasized the need to promote international cooperation in the field of AI governance, especially with LMI countries. Algorithms based on data from HI countries are often not applicable to LMI environments [23] and current regulations are often inadequate. Ethical governance of health data at the global level and regionally focused impact assessment must be prioritized [24]. To avoid a regional concentration of AI technologies in HI countries, which deprives a large part of the world, the cross-border flow of data must be ensured [67].