This scoping review follows the framework proposed by Arksey and O’Malley [19]. Briefly, the review has the following 5 stages: (1) identifying the research question, “What are the roles of AI and machine learning in the allocation of health care resources, before and during the COVID-19 pandemic?”; (2) identifying suitable studies; (3) selecting studies for review; (4) consolidating the data; and (5) summarizing and reporting the results. This study complies with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) [20] for reporting scoping review results.

Searches were conducted in MEDLINE (Ovid), PubMed, Web of Science, and Embase from inception to February 2022. The search featured 2 key terms: (1) artificial intelligence, including related terms such as big data and algorithm, and (2) health care resource allocation. The search terms were used with the “explode” feature where applicable. For example, in MEDLINE and Embase, we used exp artificial intelligence/ and exp resource allocation/, and in PubMed, relevant MeSH (Medical Subject Heading) terms were used. The search was individually designed and adapted for each database.

Inclusion and exclusion criteria were defined a priori. This scoping review includes qualitative and quantitative studies investigating the application of AI in health care resource allocation. Studies that are not relevant to AI or health care resource allocation were excluded, as were duplicate studies. The inclusion and exclusion criteria are summarized in Table 1.

Selection was conducted in 2 steps. First, titles and abstracts were screened for topic relevance and study design. Second, full texts of the remaining studies were screened to check for eligibility. All of the study selection processes were conducted in EndNote X9 (Clarivate).

Selected studies were input into NVivo 12 (QSR International) for labeling and coding. Authors coded data of interest from the articles in NVivo 12 and extracted information regarding study author, study design, location, context, aim, main result, AI method under study, resource allocation situation, and policymaking relevance into a standardized Excel (Microsoft Corp) form.

We employed an inductive approach to summarize the results from the included studies. First, the selected papers were grouped into 2 types: (1) studies of model development and validation of AI-based algorithms applied to health care resource distribution, and (2) qualitative studies, theoretical discussions, and review studies of the application of AI in health care resource distribution. For studies of model development and validation, we extracted the study objectives, resource distribution situations, AI model input variables, and policy relevance. For studies in the second category, objectives, resource distribution situations, discussed topics, and policy relevance were extracted. We further divided the input variables of the studies of model development and validation into 2 predefined categories: (1) ecological variables or variables at the group level, which included variables depicting characteristics at the population level, such as infant mortality in a region, local economic development, or disease prevalence and incidence; and (2) individual variables, which included variables that define individual characteristics such as diagnosis and age.

In total, 298 studies were identified in 4 databases after removing duplicates. After 1 round of screening for titles and abstracts, 255 studies were excluded due to irrelevant topics and unsuitable study designs. This left 43 studies for full-text screening. Of these, 2 were excluded because they were not directly relevant to health care, 8 because they were not related to resource distribution, 7 because they did not feature applications of AI, and 4 because of an inappropriate study design. In the end, 22 studies remained for qualitative synthesis. The PRISMA flow diagram for study selection is presented in Figure 1.

The characteristics of the included studies on model development are summarized in Table 2. The included studies were published between 2013 and 2021. Of the 22 included studies, 9 focused on model development and validation [21-30]. Of these, 5 studies were conducted in emerging economies, including 2 in China [27,29], 2 in Brazil [25,28], and 1 in Ecuador [26]. In developed countries, 3 studies were conducted. These included 1 in Germany [23], 1 in the United Kingdom [22], and 1 in the United States with a validation data set in China [24]. One study was applied to a global context [21].

Of the 9 studies, 4 focused on resource distribution at the health system level, including financial resources for public health in Brazil [25], health care resource distribution in health planning in Ecuador [26], medical resource allocation in the hierarchical health system in China [29], and medical equipment allocation in the global COVID-19 pandemic [21]. The remaining 5 studies focused on resource allocation at the hospital level, including bed allocation in a London hospital [22], day resources and bed allocation in a hospital in Munich, Germany [23], human resources and medical materials in a public hospital in China [27], medical resource allocation in a hospital in the capital of State of Minas Gerais in Brazil [28], and medical resource allocation in clinics for COVID-19 patients in New York [24].

The characteristics of studies involving reviews and theoretical discussions are summarized in Table 3. Of the 22 included studies, 13 were theoretical discussions, qualitative studies, or review studies [31-43]. Of those studies, 8 studies were qualitative discussions on the COVID-19 pandemic [31,33,34,36,38,39,41,43], with 2 in a Chinese context [34,43] and the rest in a global situation. The remaining 5 studies focused on other situations, with 1 focusing on resource allocation in intensive care units and hospital stay [40], 1 on disease outbreaks and disasters [33], 1 on diabetic retinopathy screening [42], 1 on human resource allocation in health systems [35], and 1 on medical information digitalization [37].

The policy implications of studies on model development are relevant on 2 levels: (1) health system level [21,25,26,29] and (2) hospital level [22-24,27,28], corresponding to situations where the models were applied. Detailed policy implications of the included studies on model development are summarized in Table 4. The qualitative and review studies focused largely on 2 issues: (1) how AI can promote the efficacy of resource allocation [21,32,34-37,39,42,43] and (2) the ethics and equality issues associated with using AI systems [38,40,41]. One study highlighted the lack of AI studies on resource distribution during COVID-19 [31]. Table 5 summarizes the policy implications of these studies.

China and Brazil are both developing countries with a similar per capita gross domestic product (China: US $10,435 and Brazil: US $6797) [44]. During the COVID-19 pandemic, Brazil has had one of the highest national overall cases and mortalities, as well as per capita cases and mortalities, with 29.5 million cases and 656,000 deaths as of March 2022 [45]. China has had one of the lowest per capita infection rates in the world, with a total of 124,000 cases and 4636 deaths as of March 2022 [45]. Given the similarity between the 2 countries in economic development and the enormous difference in COVID cases and mortalities, the resource distribution situation in the 2 countries is worth exploring.

Rosas et al [25] proposed a financial resource allocation algorithm for the public hospital system in Brazil based on mortality, socioeconomic characteristics, and income inequality. They argued that the choice of input variables for health care policymaking should consider the vulnerability of the population to being manipulated by those who manage public policy, the true representation of the factors of need, exemption from the process of political choice, and the availability of reliable data. The focus of the model was regional economic characteristics.

Zhang et al [29] proposed a model for the allocation of medical resources and tier classification of patients in China’s health system, with the input variables of patient characteristics and hospital tiers, and a focus on differentiation into different tiers based on patients’ disease severity. Xu et al [27] proposed a health resource allocation model for the allocation of doctors and other medical resources in a public hospital system in China that considered the distribution of medical stations, the professional level of doctors (salary and seniority), patient preferences and illness severity, medical cost, and revenue.

Overall, the allocation of medical resources based on the models from the 3 studies demonstrated that the key considerations proposed by studies from China were the hospital tier system, the professional level of doctors, the geographical distribution of medical resources, and cost-effectiveness [27,29]. However, the model proposed for Brazil focused on the regional economic situation [25].

In this review, we compiled evidence on the application of AI in health resource distribution, especially regarding COVID-related policy. After synthesizing 22 articles, we found that AI-based models were proposed at both hospital (secondary care in inpatient settings) and health system (public health) levels and that theoretical discussions and reviews focused on the potential for AI to improve the efficacy of resource distribution and on the ethics of applying AI in health resource distribution. Two major themes emerged from the review. First, we found that AI-informed resource distribution strategies are impactful for health access and equality. Second, the approaches can be categorized ideologically into revisionist and conservative groups.

AI and machine learning have considerable potential to improve efficacy in resource distribution, especially during emergencies, such as the COVID-19 pandemic, where quick decisions are required based on evolving situations [34,39,43]. For example, health informatization, particularly digital contact tracing and AI-informed response design, played an instrumental role in responding to COVID in China and helped local governments to improve efficacy in allocating limited resources [34,43]. AI can also be used to interpret diagnostic results and patient characteristics in order to predict disease progression and allocation of medicines, hospital beds, and medical professionals at the hospital level [21,39].

However, very large amounts of data are necessary for AI algorithms to make reliable and evidence-based decisions [46]. Health care institutions globally must therefore collect, record, and analyze data. This will help policymakers gather novel insights and translate the data into a prompt, equal, coordinated, and more successful response to the next pandemic [32,47]. As such, data collection must be institutionalized. The disparity in data collection capacity potentially exacerbates the gap in decision-making quality between countries [48,49]. For example, from the literature, China’s information infrastructure and data-sharing agreements expedited the data-gathering process, a possible consequence of the centralized government system that facilitated gathering data, which in turn made the data set larger and more comprehensive [48]. In contrast, a selected study showed that Brazil’s decentralized government system, with heterogeneous policies on data privacy and data sharing, made the collection and consolidation of data difficult [49]. However, caution should be taken in interpreting those results, as there is no evidence that the studies selected here are representative of the real situation in China or Brazil.

The included articles highlighted the importance of distributive justice and transparency in AI model design. The analysis conducted by Rajkomar et al emphasized that machine learning systems should be used proactively to advance health equality [40]. They proposed that distributive justice should be a core principle in AI models, including during the design, deployment, and evaluation processes. This perspective would include equality in patient outcomes, performance for every sociodemographic group, and resource allocation for each group. As Neves et al noted, resource allocation by AI and in emergencies should build on basic ethical values, including the equal value of people, instrumental value, and priority for critical situations. Transparency is the key to gaining trust when distributing resources [38].

The build-up of AI models and implementation plans can be broadly categorized into revisionist and conservative approaches. In revisionist approaches, the models aim to revise the disparity in resource distribution by actively correcting the biases in previous decision-making processes. For example, the models proposed by Rosas et al [25] for financial resource allocation in Brazil emphasized consideration of income inequality, vulnerable populations, political choices, and the availability of reliable data. In conservative approaches, the models rely on traditional metrics, including supply and demand, profitability, and, perhaps most notably, previous decisions. This was demonstrated in a proposed model for the allocation of medical resources and tier classification of patients in China’s health system by Zhang et al [29], where the input variables were patients’ characteristics and hospital tiers, and a model suggested by Xu et al [27] for the allocation of doctors and other medical resources in a public hospital system in China, where the input variables included the distribution of medical stations, the professional level of doctors, patient preferences and illness severity, medical cost, and revenue. Doctor expertise, patient characteristics, hospital tier, and location are common variables in human decision-making, but AI has the potential to analyze the data more thoroughly.

However, despite the revisionist model proposed by Brazilian academics [25], health inequality is a prevailing issue in Brazil across states and social classes, both before [50] and during the COVID-19 pandemic [51]. Health inequality in Brazil increased across states from 1990 to 2016 [43]. Comparatively, the health care access and quality index in China was higher than that in Brazil in 2016, suggesting better equality and health care access in China [52]. However, due to the limitation of the research method, this study could not show the policymaking processes in both countries. From the selected studies alone, we observed that although proposing revisionist AI models to address health inequality should be encouraged, the application and practicality of using those models to inform health policy decisions and improve inequality should also be important considerations for researchers.

This is one of the first reviews to incorporate all available evidence qualitatively and provide a comprehensive picture of the model development and theoretical discussion on AI in medical resource distribution. Our results contribute to the ongoing discussion of applying AI in medical resource distribution and add novel insights into the social and ethical implications. Nonetheless, this study has several limitations. First, due to the scope of the study, we focused on published journal articles but did not examine policy documents or grey literature. This could have led to incompleteness in the collected information. Further studies could examine policy statements and grey literature to better understand intercountry differences. Second, we included only articles published in English and therefore might have overlooked publications in other languages. Third, there are potential sources of meaningful heterogeneity in this scoping review, including the diverse use of AI technologies, different study designs, and different locations. The analyses in this study could be affected by such heterogeneities. Fourth, this study is a qualitative overview of the general application of AI in health care resource distribution and is exploratory. We did not compare different levels of resource distribution and distinguish various machine learning methods in detail. Further studies are needed to explore and contrast different AI approaches at various resource distribution levels in detail. Lastly, due to the availability of evidence, we only compared studies from China and Brazil. We were only able to compare the differences between the 2 countries based on a few studies, which could not represent the real situation in either country. The comparison should be interpreted as exploratory and demonstrative.

This scoping review synthesized evidence on the application of AI in health resource distribution, particularly during the COVID pandemic. The included studies suggested that AI and machine learning have high potentials to improve efficacy in resource distribution, especially during sudden and evolving situations. A coordinated and continuous data sharing and collecting mechanism is needed for better data input so that AI can make reliable and evidence-based decisions. Various issues, including health inequality, distributive justice, and transparency, should be considered when deploying AI models. Such considerations are required for implementing revisionist AI models that can correct distribution inequality in actual policy processes.