The widespread and rapid adoption of artificial intelligence (AI) technologies in health sciences, education, and practices introduces new ways of delivering patient care [1]. AI encompasses a broader term within computer science, which includes technologies that can incorporate human-like perception, intelligence, and problem-solving into complex machines [2]. Big data in health care, along with high-performance computing power, has enabled the use of AI, machine learning (ML), and deep learning, in particular, to improve clinical decision-making and health sector efficiency [3]. More recently, AI-enabled technologies have continued to emerge, predominantly in the medical fields of radiology, anesthesiology, dermatology, surgery, and pharmacy [4-7]. Although AI is not likely to replace clinical reasoning, Mesko [8] predicts that AI will influence all specialties in varying degrees, depending on the nature of the practice (eg, the degree of repetitive tasks involved and whether the tasks are data driven). However, the efficacy of AI-enabled technologies in health care depends on the involvement of health care professionals (HCPs) in developing and validating these technologies. Therefore, HCPs should play a role in this transformation and be involved in every aspect of shaping how AI adoption will affect their specialties and organizations.

Recommendations for HCP involvement are emerging. For instance, in the field of medical imaging, West and Allen [9] recommend that HCPs be involved in (1) implementing data standards and following them in practice, (2) prioritizing use cases of AI in medicine, (3) determining the clinical impact of potential algorithms, (4) describing and articulating the needs of the profession for data scientists and researchers, and (5) participating in the translation of practice needs from human language into machine language. As these technologies emerge, it is essential for HCPs and educators to have the competencies required to rapidly develop and incorporate these changes into their practices and disciplines.

At the individual level, a lack of AI literacy is a significant barrier to the adoption and use of AI-enabled technologies to their full capacity in various medical specialties. In AI education programs specifically, there are barriers to implementation at various levels of medical education (undergraduate, postgraduate, practice-based education, or continuing professional development). For instance, health informatics plays a valuable role in modern medicine; yet, it is not the focus of most medical school curricula [2]. Technology experts are often consulted to provide training on the use of electronic clinical tools, but this does not support the level of skill required to understand how it could be used to enhance patient interactions and improve care [10]. Another example exists within radiology residency programs, where the lack of awareness as well as lack of knowledge of implementing and using AI were cited as barriers to its adoption [11,12]. Incorporating AI fundamentals into health professionals’ curricula is essential, and it would be useful to balance this knowledge with providing patient-centered care by empowering future HCPs to consider AI in the context of their own clinical judgment. The combination of trust in their own judgment and basic statistical knowledge will be useful in understanding how to best apply new AI-driven technologies in clinical practice [13]. AI needs to be considered within the context of HCPs’ broader skill sets, priorities, and ultimate goals in health care; this includes encouraging patient-centered, compassionate care in clinical practice [13,14].

Martec’s Law refers to the idea that technology changes occur much more rapidly, and in fact exponentially, compared with the ability of organizations to adopt these technologies [1]. Therefore, organizations need to promote innovative technologies proactively and empower their professionals to be adequately trained to successfully implement AI-based tools in their practice [1]. A concerted, deliberate approach is required to incorporate these new technologies, both effectively and compassionately, at an individual level and within the culture and operations of an organization [1].

A number of potential barriers to implementing these technologies exist; the 3 main limitations identified include regulatory, economic, and organizational culture issues [15]. Regulatory approval [16] is needed to adopt AI technologies in clinical settings, and potential liabilities in using these technologies for patient care must be considered, as well as the safety, efficacy, and transparency of AI algorithms for clinical decision-making [17,18]. Regulatory issues can also come into play when it comes to accessing data for AI adoption; multi-institution data sharing is required for algorithm improvement and validation, as well as the accompanying research ethics board and regulatory approvals [18]. To further improve adoption, these technologies will also have to be economical, supported by adequate funding [18], and seem as valuable to the organization itself. At an organizational level, the use of AI should align with the goals and strategic plans of an organization; organizations will need to assess how well the AI technology will integrate into existing systems, including data warehouses and electronic health records [18]. It may be difficult to generalize a particular AI model across different clinical contexts to a degree that would prove valuable at an organizational level while still working seamlessly and being clinically useful at the individual level [15]. Furthermore, when choosing to adopt AI technologies, organizations can either collaborate with outside vendors or create the technologies in-house, which will require the use of additional human and material resources [15].

Deficits in AI education may be contributing to a lack of capacity in health care systems to fully integrate and adopt AI technologies to improve patient care, despite calls for AI integration as part of the National Academy of Medicine’s Quintuple Aim Model [19]. It is important to equip health care organizations and their stakeholders to have the cognitive, psychomotor, and affective skills to harness AI in enhancing and optimizing the delivery of care. This will also involve supporting AI education initiatives that are widely available for all types of HCPs. To support future AI education development, dissemination, and evaluation, it is important to assess the current situation within AI adoption in health care and further understand the extent of AI education implementation, including who is receiving AI training or education, what content is covered, how it is delivered, and whether this reflects what experts believe that AI education curricula should include. Therefore, this scoping review aims to establish a foundational understanding of education programs on AI for HCPs by determining the following:

This scoping review followed the Arksey and O’Malley [22] guidelines and the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) Extension for Scoping Reviews checklist [23,24]. The objective of this scoping review is to examine and summarize the extant literature on AI education and training for HCPs.

A health sciences librarian (MA) developed strategies for Ovid MEDLINE All, Ovid Embase, Ovid APA PsycINFO, Ovid Emcare Nursing, Ovid Cochrane Database of Systematic Reviews, Ovid Cochrane Central Register of Controlled Trials, EBSCO ERIC, and Clarivate Web of Science using appropriate subject headings and keywords for AI and health professions education. As a result of the widespread use of terms relating to health professions and education in health sciences literature, the decision was made to focus the searches on health professions education concepts to reduce noise in the results sets. Searches for these subject headings were limited to where they were the major subject heading (the most important subject heading in the database record for an item). Keywords for these concepts were only searched in the study titles, the author-assigned keywords, heading words, and journal titles, depending on the content and field availability of the database. No language or date limits were applied. The searches were run and the results were downloaded on July 7, 2020. For the complete strategies, see Multimedia Appendix 1. If the search results included conference abstracts and proceedings, a subsequent search to find any corresponding follow-up studies was conducted in Google Scholar. Finally, pearl growing, also known as a hand comb process, was conducted where all cited works in the included studies from the initial screening underwent a 2-stage screening process (title and abstract scan as well as full-text review).

The 2-stage screening process consisted of (1) title and abstract scan and (2) full-text review. Study eligibility was determined by 2 independent reviewers, and a third reviewer was involved to resolve any conflict when consensus was not reached between the 2 reviewers. For a study to be included for full-text review and to be chosen for subsequent inclusion, the title and abstract at each stage needed to have the following attributes:

A pilot review of 20% (595/2973) of the MEDLINE citations was conducted to establish interrater reliability. The interrater reliability threshold had a Cohen κ value of 0.70, indicating substantial agreement. Additional batches of 50 citations were reviewed until the threshold was met.

A standardized charting form was developed to capture the following domains: article details, study details (if publication was an empirical study), education program details, and implementation factors. The subdivisions of the domains for the data extraction are outlined in Table 1.

To collate, summarize, and report on the included studies in this review, a narrative synthesis approach was used [25]. This included a numeric summary using descriptive statistics to report each domain (article details, study details, education program details, and implementation factors). For program curriculum under education program details, curriculum topics were inductively coded. Once a list of topics was generated, they were then grouped by domain using the taxonomy for learning formulated by Bloom. There are 3 domains: (1) cognitive, which refers to knowledge that learners should have, (2) psychomotor, which refers to skills learners should demonstrate and master, and (3) affective, which refers to attitudes learners should develop and incorporate into their practice [20].The study outcomes were deductively coded using the Kirkpatrick-Barr Framework [21] of educational outcomes. This framework was selected because it provided a standardized method of categorizing the type of educational outcomes reported by each study. The implementation factors subdomain was thematically analyzed by 2 independent reviewers using a priori codes. The reviewers compared coding schemes and iteratively determined overarching themes to frame their findings. For content validation, the project team members, patients, and experts in the fields of medical education and AI provided feedback on the thematic analysis.

The initial database search yielded 13,449 results; once duplicates were removed, the titles and abstracts of 10,094 (75.05%) unique citations were identified. From the 10,094 unique citations, we identified 41 (0.41%) articles [2,5,13,26-63], where 13 unique, existing programs [32,35,39,43,49,50,59,61-63] were mentioned in 10 (24%) articles, and the remaining 31 (76%) articles [2,5,13,26-31,33,34,36-38,40-42,44-48,51-58,60] discussed the desired or recommended curricular content. The article selection process is presented in Figure 1. Of the 10 articles that discussed an existing program, 8 (80%) were commentaries [32,43,49,50,59,61-63], 1 (10%) was a case report [39], and 1 (10%) was an empirical study [35]. Tables 2 and 3 describe the characteristics of the articles and programs included in this review.

Summaries of the individual programs can be found in Table 4. Of the 13 programs, 8 (62%) originated from the United States [32,35,49,50,59,61-63], 1 (8%) from Canada [43], 1 (8%) from France [59], and 1 (8%) from Mexico [39]. The typology described by Strosahl [64] was used to classify the educational method. Of the 13 programs, 9 (69%) had a didactic approach [32,35,39,43,50,59,61,62] in combination with discussions [62] (1/13, 8%), web-based [39,61] (2/13, 15%), workshop and case-based [50] (1/13, 8%), and experiential learning [43] (1/13, 8%). Of the 13 programs, 10 (77%) were taught in an academic setting [32,35,39,43,59,61,62].

There were 3 types of HCPs identified in the 41 reviewed papers: physicians [41,43,46,52,59,63] (6/41, 15%), nurses [31,52] (2/41, 5%), and radiology technologists [5] (1/41, 2%). In addition, 2 specific specialties were identified: medical imaging [5,26,32-34,37,42,48,50,51,55,58,63] (13/41, 32%) and cardiology [56,61] (2/41, 5%), with others not being specified [2,13,27-31,35,36,38-41,43-47,49,52-54,57,59,60,62] (26/41, 63%). Figure 2 illustrates the type of curriculum topics covered in the continuum of learning for clinicians, which includes undergraduate medical education [2,13,28-30,33,35-37,39-41,43,44,47,53,57,61,62] (20/41, 49%), postgraduate medical education [5,26,32-35,41,42,48-51, 54-58,62,63] (19/41, 46%), and continuing professional development [5,57] (2/41, 5%). Other nonclinical professionals include researchers [5,27,33,59,62] (5/41, 12%), health care administrators [27,33,45,52,62] (5/41, 12%), and computer and data scientists [27,33,52,63] (4/41, 10%).

From these papers, the program curriculum and desired or recommended content mentioned included topics on using AI, interpreting AI, and explaining results from AI, as framed by McCoy et al [43]. A description of each curricular topic can be found in Table 5.

Of these 16 curricular topics, 9 (56%) fell under cognitive domain, 6 (38%) under psychomotor domain, and 1 (6%) under affective domain, and most of them were mentioned both by papers that discussed current education programs and commentaries that discussed what HCPs should be learning. The curricular topics were categorized into the 3 domains identified in the taxonomy for learning formulated by Bloom [20]. Table 6 displays the curricular topics that were unique to 24% (10/41) of the papers [32,35,39,43,49,50,59,61-63] that described what AI programs currently teach, 76% (31/41) of the papers [2,5,13,26-31,33,34,36-38,40-42,44-48,51-58,60] that described what AI programs should teach as part of their curriculum, and those that outlined both what was taught and what should be taught.

Of the 41 papers, 20 (49%) [5,26,32,34,36,37,39-41,43, 45-47,50,52,55,57-59,62] highlighted the importance of providing HCPs with a baseline understanding of AI and 10 (24%) [32,35,39,43,49,50,59,61-63] recommended teaching them AI applications. The studies focused on various applications of AI, including diagnostic systems, data gathering, assessment and use, clinical applications, and personalized care. In addition, many of the papers reported that medical curricula should integrate fundamentals of health care data science [5,13,26-36,42,45,49,50,52,54] (19/41, 46%), including, but not limited to, big data and bioinformatics. Matheny et al [45] stated that data science curricula should encompass how to form multidisciplinary development teams to improve the value of AI and to be aware of the ethics, equity, diversity, and inclusion principles at play and the inadvertent ramifications that may result from AI implementation. The studies also focused on statistics, ML with model development, model translation and use in clinical knowledge, data extraction, and applications for visualization of patients. Familiarity with ML vocabulary and a basic understanding of the methodology (algorithms and machine gathering and process of data) were deemed important to understand this rapidly emerging field.

Most of the papers focused on clinicians being able to effectively analyze the data [2,5,31,34,35,40,43,46,50,54,55,58,59,61,63] (15/41, 37%) to identify trends and efficiency correlations. As highlighted by Balthazar et al [63] and Forney and McBride [55], it is imperative to learn how to evaluate the efficacy and precision of AI applications. This point was reinforced in a review conducted by Park et al [40] that stated medical students should be able to validate the clinical accuracy of algorithms. HCPs will need to become accustomed and understand how to embrace real-time health information to help make decisions in their practice setting [61]. Of the 41 papers, 8 (20%) discussed the significance of understanding and interpreting the findings with a reasonable degree of accuracy, including awareness of source error, bias, or clinical irrelevance [13,28-31,39,47,50]. Moreover, the study findings described problem-solving [35,38,60] (3/41, 7%) as a critical skill, entailing the management and application of several distinct resources. Clinicians will need to become adept in communicating the results and processes [5,28-30,32,36,43,46] (8/41, 20%) with patients in a personalized and meaningful manner. Cultivating and expressing empathy and compassion [28-30,36] (4/41, 10%) when communicating with the patient was emphasized in several studies.

Of the 41 papers, 8 (20%) stressed that HCPs should have the attitude to harness AI tools effectively to improve outcomes for patients and their communities [27,30,37,52,55,58,59,62]. Wiljer and Hakim [27] asserted the importance of breaking the mass stereotypes about AI as an initial step. It is essential that professionals perceive AI as augmenting their delivery of care, rather than taking over different aspects of the health care system [62]. Forney and McBride [55] stated that clinicians are not as likely to perceive AI as a threat if they are able to see the wide array of AI tools and the impact these tools have on workflow and patient care. Furthermore, Sit et al [37] mentioned that medical students are not as likely to be discouraged from pursuing certain specialties when they are presented with use cases and understand the boundaries of AI tools; almost half of the respondents believed the misconception that because of AI, certain specialists such as radiologists will become obsolete in the near future. Moreover, Brouillette [59] mentioned the need for collaborative programs among medical students, computer science students, and engineering students, where they can better understand each other’s disciplines. A few papers recommended that future AI programs should integrate change management and establish a culture of trust and transparency with relevant stakeholders, which will support organizations to more rapidly adopt and implement AI technologies within the health care ecosystem [27,30]. Thus, it is vital to help organizations manage change at a rate in pace with the rapid advancement of technology.

Of the 41 papers, 5 (12%) presented the results of their training evaluation [35,39,49,50,62]. As the educational approaches varied across studies, each approach will be briefly discussed (Table 7), followed by the measures and outcomes associated with each educational initiative. Categorized according to the Kirkpatrick-Barr Framework, the outcomes were either level 1 (ie, learner reaction and satisfaction with the education) [39,49,50], level 2a (ie, change in attitude) [49,50,62], or level 2b (ie, change in knowledge or skill) [35,62]. There were no outcomes that could be categorized as level 3 or level 4; thus, the program evaluations did not comment on the change in behavior or affect at the organizational level or on patient outcomes.

This review identified pivotal knowledge gaps in our understanding of effective AI education programs for HCPs. The gaps identified through this review illustrated the limited AI education and training opportunities available for HCPs and thus emphasized the necessity of curating further AI education programs targeted to HCPs. The existing programs tend to focus only on the development and implementation of AI; yet, it is essential to also prepare HCPs to not only work with AI but also to advance AI for health and clinical decision-making. AI education programs should be designed in a way that enables HCPs to not only safely adopt these technologies, but also to adapt and shift their scope of practice to stay relevant. A significant and meaningful change to AI curricula in health care will only occur by increasing AI literacy among HCPs and by providing them with the ability to leverage relevant digital and data-driven decision-making tools. Although the studies demonstrate that efforts are being made to evaluate the outcomes of AI-related education initiatives, there is a lack of consistency in the measures for a comprehensive assessment of these outcomes. Most of the papers used self-constructed and nonvalidated instruments and delineated their findings in qualitative terms. Given the variety of instruments that have been employed in the studies, the absence of a standard, comprehensive tool impedes the integration and synthesis of findings across the studies. The guiding principles provided in this review will also hopefully inform future development and design of these programs.

A lack of infrastructure to integrate AI content into current curricula could hinder the development of these types of programs; some of the programs described embedded their content within existing professional certifying bodies’ infrastructure to facilitate content development. The Royal College of Physicians and Surgeons of Canada, in particular, further emphasizes the need for these regulatory strategies, which are currently in process but not yet in practice [65,66]. The promotion of multidisciplinary collaboration was indicated as an enabler of content delivery; yet, varying levels of AI literacy among faculty could impede successful delivery of AI content [54,57]. Curricular adaptations and building an infrastructure for AI technologies could be helpful to HCPs wanting to adopt AI to improve patient care; this includes improvements in the types of health care data available for AI education [67]. Of note, much of the health data generated are often inaccessible to researchers and limited by regulatory or infrastructure-level barriers, including institutional ethics approvals and data-sharing agreements [67]. The use of deidentified data, security, and privacy controls could potentially widen the scope of access; broader collaboration with multidisciplinary experts could also help to establish secure data networks to improve use and access of health care data [67]. Lower levels of AI literacy could be augmented by standardizing competency statements and engaging and training faculty in e-learning, for instance [68-70]. The World Health Organization’s Global Strategy on Digital Health further suggests that the barriers to AI adoption need to be addressed at the systems level and all aspects of implementation should be considered.

Our recommendations have been formed into guiding principles that could be used to guide the development of future AI curricula or to incorporate AI education into existing curricula.

Our scoping review findings should be examined in the context of the following limitations. Because of the nature of the scoping review, the quality of each identified study was not assessed. Given the nature of the topic being investigated, we excluded studies that discussed AI as a tool for medical education or continuing professional development. Only studies in English were included. In addition, the educational approaches varied across the studies; thus, we were unable to conduct formal comparisons among the curricula to determine which were effective. However, reviewing the literature enabled us to identify the gaps in current education programs and provide insights and best practices to guide future education efforts. As this review was inclusive of all types of studies and focused on a breadth of literature, the depth in reporting of education program details was inconsistent and varied based on the scope of the study.

With the inevitable progression of health care digitization, health professions education should foster unique human abilities, which will complement these emerging technologies. This review provided an overview of the current state of AI in health professions education and future directions on preparing care providers for the era of AI in health care. Future education efforts should focus on the development of regulatory strategies, a multidisciplinary approach to curriculum redesign, a competency-based curriculum, and patient-clinician interaction.