Perceptions and Needs of Artificial Intelligence in Health Care to Increase Adoption: Scoping Review

Background: Artificial intelligence (AI) has the potential to improve the efficiency and effectiveness of health care service delivery. However, the perceptions and needs of such systems remain elusive, hindering efforts to promote AI adoption in health care. Objective: This study aims to provide an overview of the perceptions and needs of AI to increase its adoption in health care. Methods: A systematic scoping review was conducted according to the 5-stage framework by Arksey and O’Malley. Articles that described the perceptions and needs of AI in health care were searched across nine databases: ACM Library, CINAHL, Cochrane Central, Embase, IEEE Xplore, PsycINFO, PubMed, Scopus, and Web of Science for studies that were published from inception until June 21, 2021. Articles that were not specific to AI, not research studies, and not written in English were omitted. Results: Of the 3666 articles retrieved, 26 (0.71%) were eligible and included in this review. The mean age of the participants ranged from 30 to 72.6 years, the proportion of men ranged from 0% to 73.4%, and the sample sizes for primary studies ranged from 11 to 2780. The perceptions and needs of various populations in the use of AI were identified for general, primary, and community health care; chronic diseases self-management and self-diagnosis; mental health; and diagnostic procedures. The use of AI was perceived to be positive because of its availability, ease of use, and potential to improve efficiency and reduce the cost of health care service delivery. However, concerns were raised regarding the lack of trust in data privacy, patient safety, technological maturity, and the possibility of full automation. Suggestions for improving the adoption of AI in health care were highlighted: enhancing personalization and customizability; enhancing empathy and personification of AI-enabled chatbots and avatars; enhancing user experience, design, and interconnectedness with other devices; and educating the public on AI capabilities. Several corresponding mitigation strategies were also identified in this study. Conclusions: The perceptions and needs of AI in its use in health care are crucial in improving its adoption by various stakeholders. Future studies and implementations should consider the points highlighted in this study to enhance the acceptability and adoption of AI in health care. This would facilitate an increase in the effectiveness and efficiency of health care service delivery to improve patient outcomes and satisfaction. (J Med Internet Res 2022


Introduction
Background Rapid advances in artificial intelligence (AI)-software systems designed to mimic human intelligence or cognitive functions-have sparked confidence in its potential to enhance the efficiency of health care service delivery and patient outcomes [1][2][3]. However, although AI has been rapidly adopted in many industries, such as finance and information technology (IT), its adoption in health care remains relatively lagged because of the ethical and safety considerations that are more pronounced when it comes to human lives at stake [4]. AI-powered systems in health care can autonomously or semiautonomously perform a wide variety of tasks, such as medical diagnosis [5], treatment [6], and self-monitoring and coaching [7,8]. In some studies, AI has been shown to outperform human capabilities, such as analyses of chest x-ray images by radiologists [9]. Not only is AI expected to improve the quality of care and health outcomes for patients by decreasing human errors, but it is also likely to free up time for clinicians and health care workers from routine and repetitive tasks, enabling them to focus on more complex tasks [9,10]. For instance, in many areas of medical imaging, the use of fast and accurate AI-assisted diagnoses would significantly increase the workflow efficiency by processing more than 250 million images per day [11]. Various AI chatbots have also been developed to provide mental health counseling and assist overburdened clinicians [9]. Through AI-enabled apps and wearable devices, patients and the public could self-monitor and self-diagnose symptoms, such as atrial fibrillation, skin lesions, and retinal diseases [9].
Owing to the emerging nature of modern AI systems, the perceptions and needs of affected stakeholders (eg, health care providers, patients, caregivers, policy makers, and IT technicians) on the use of AI in health care are not yet fully understood. A large body of literature suggests that human factors, such as trust, perceived usefulness, and privacy, play an important role in the acceptance and adoption of past technologies in health care, including handheld devices [12], IT [13], and assistive technologies [14]. However, current evidence remains broad and general, and little is known about the perceptions and needs of AI in community health care. As the world makes a paradigm shift from curative to preventive medicine, AI holds a strong transformative potential to enhance sustainable health care by empowering self-care, such as self-monitoring and self-diagnosis. However, it is important to first understand the perspectives of all direct users of AI-driven systems (eg, patients and frontline health workers) and their perceived needs to ensure its successful adoption across different parts of the health care sector, especially community health care. Thus, this study aims to present an overview of the perceptions and needs of AI in community health care. The implications of this study will help inform the design of future health care-related AI technology to better fit the needs of users and enhance the adoption and acceptability of the technology.

Definition of AI
First, as the term AI is broadly used in many disciplines to represent various forms of intelligent systems and algorithms, it is important to establish a concrete and unified definition of AI for this study. Specifically, we adopted the definition of AI proposed by the High-Level Expert Group on Artificial Intelligence [15], which describes AI in terms of both a technology and field of study: Artificial intelligence (AI) systems are software (and possibly also hardware) systems designed by humans that, given a complex goal, act in the physical or digital dimension by perceiving their environment through data acquisition, interpreting the collected structured or unstructured data, reasoning on the knowledge, or processing the information, derived from this data and deciding the best action(s) to take to achieve the given goal. AI systems can either use symbolic rules or learn a numeric model, and they can also adapt their behaviour by analysing how the environment is affected by their previous actions. Furthermore, most, if not all, modern AI systems are considered artificial narrow intelligence (ANI) or Weak AI [15] designed to perform one or more specific tasks. In health care, domain-specific tasks for ANI may vary from performing human perceptions, such as image recognition [16] and natural language processing [17], to making complex clinical decisions, such as medical diagnostics [18]. Many recent advances and breakthroughs in ANI use learning-based approaches, namely, deep learning, in which computational models consisting of several layers of artificial neural networks (hence the titular deep) are trained by learning from a massive amount of sample data to perform specific tasks. Although recent performances of ANI appear very promising, ANI models are limited in their generalizability, that is, models trained to perform tasks in one domain cannot be generalized to other domains. For example, ANI trained to diagnose diabetic retinopathy from fundus images cannot be directly used to detect pneumonia from chest x-ray images. In contrast to ANI, artificial general intelligence (AGI) or Strong AI [15] belongs to a class of AI that displays true human intelligence, capable of continuously learning and performing any tasks like a real human. AGI is most likely in public consciousness when talking about AI, as it is frequently portrayed in popular culture by sentient robots and self-aware systems. At present, no AI systems have been able to come close to exhibit the AGI capability. For a useful and concise summary regarding the definitions, terminologies, and history of AI, see the following technical reports: Ethics Guidelines for Trustworthy AI [15] and Historical Evolution of Artificial Intelligence [19].

Methods
A systematic scoping review was conducted according to the 5-stage framework by Arksey and O'Malley [20]. Results were reported according to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) checklist (Multimedia Appendix 1) [21].

Stage 1: Identifying the Research Question
Our research question was as follows: What is known about the perceptions and needs of AI in health care?

Stage 2: Identifying Relevant Studies
Studies were searched from inception until June 21, 2021, using a 3-step search strategy. First, potential keywords and Medical Subject Headings terms were generated through iterative searches on PubMed and Embase. Keywords such as machine learning did not result in better search outcomes (ie, many irrelevant results were retrieved, such as the use of machine learning to explore perceptions of other topics); hence, they were omitted. Next, keywords including artificial intelligence, AI; public; consumer; community; perception*; preference*; needs*; opinions*; and acceptability were searched through nine databases: ACM Library, CINAHL, Cochrane Central, Embase, IEEE Xplore, PsycINFO, PubMed, Scopus, and Web of Science. Additional articles were also retrieved from the first 10 pages of the Google Scholar search results and the reference lists of the included full-text articles. The specific database searches combined with Boolean operators are detailed in Multimedia Appendix 2.

Stage 3: Study Selection
After removing duplicate articles, titles and abstracts were first screened by HSJC for inclusion eligibility. Articles were included if they were (1) focused on the use of AI in health care, except those focused on using AI to improve surgical techniques; (2) focused on perceptions, needs, and acceptability of AI in health care; (3) empirical studies or systematic reviews; (4) on adults aged ≥18 years; and (5) used in a community setting.
Articles were excluded if they were (1) not specific to AI (eg, general eHealth or mobile health); (2) pilot studies, commentaries, perspectives, or opinion papers; and (3) not presented in the English language. In total, 43 full-text articles were screened independently by both coauthors, and discrepancies were resolved through discussions and consensus.

Stage 4: Charting the Data
Data were extracted by HSJC using Microsoft Excel according to the following headings: author, year, title, aim, type of publication, study design, country, AI applications in health care, data collection method, population characteristics, sample size, age (mean or range), proportion of men, acceptability, perceptions, needs and preferences, and limitations.

Overview
Several positive perceptions on the use of AI in health care were highlighted in our findings ( Table 2). Could support the self-care needs of older people-mobility, self-care and domestic life, social life and relationships, psycholog-Able to collect data nonintrusively Abdi et al [28] mercially ready to support the care needs of older people ical support, and access to health care; potential uses for remote monitoring and prompting daily reminders, for example, medications  [31] to treat a chat-over data sharing bot as a real physician or an adviser Less than half of the social NS Social media users were pes-Distrust of AI companies ac-NS NS NS Gao et al [22] media posts ex-simistic about counted for a pressed that AI the immaturity quarter of all would complete-of AI technology negative opinions among social media users ly or partially replace human doctors NS NS There were concerns with chat-There were concerns with chat-

NS
The majority were interested in using a NS Griffin et al [25] bots making bots providing chatbot to help manoverwhelming too much infor-age medications, redemands for mation and invading privacy

Availability and Ease of Use
Of the 26 studies, 3 (12%) studies highlighted the advantage of AI being constantly available without restrictions such as physical location, time, and access to a structured treatment [24,30,38]; 3 (12%) other studies also mentioned the appreciation of respondents for how an AI system could collect data remotely in a nonintrusive and user-friendly manner [23,24,28]. These studies mostly represented the perceptions of consumers and health care providers [24,30,38] (Multimedia Appendix 3). Only 4% (1/26) of studies did not mention the population characteristics [24].

Improves Efficiency and Reduces the Cost of Health Care Service Delivery
In all, 58% (15/26) of studies highlighted the potential of AI to improve the efficiency of health care service delivery in terms of remote monitoring [28], providing health-related reminders [23,28], increasing the speed and accuracy of health care processes (eg, consultation wait time, triaging, diagnosis, and managing medication refills) [26,29,30,[35][36][37]44], facilitating care team communications, improving care accountability (eg, regular check-ins and follow-ups for information gathering) [23], and taking over repetitive manual tasks (eg, scheduling, patient education, and vital signs monitoring) [27]. Some respondents also appreciated the use of AI to provide a second opinion to physicians' diagnoses or evaluations [42,46]. Overall, 12% (3/26) of studies [24,34,45] discussed the potential cost-saving capacity of AI that influences AI acceptability, whereas 4% (1/26) mentioned that the provision of an AI service using IBM Watson caused patients to incur higher treatment costs that did not translate to profits for the hospital after factoring onboarding of the technology [40]. There was a good proportion of representation from the health care and IT staff (53.3%) [27][28][29]36,37,40,42,44] and those from the public, including patients (Multimedia Appendix 3). Only 4% (1/26) of the studies did not mention the population characteristics [24].

Overview
Our findings highlight several concerns ( Table 2) and mitigation strategies (Table 3). NS Interaction was too long, the use of nonverbal expressions by the avatar was not appealing, and there was a lack of clarity regarding the aim of the chatbot. Better integration of the agent with electronic health record systems (for a virtual physician) or health care providers (for an asthma self-management chatbot) would be useful Need for greater interactivity or relational skills in conversational agents. Respondents liked that the agent had a personality and showed empathy, which improves personal connection. Others had difficulty in empathizing with the agent or reported disliking its limited conversation and responses Need more customization or availability of feature options (eg, preformatted or free-text options) NS Milne-Ives et al [23] There was a general lack of familiarity and understanding of health chatbots among participants NS Lack of empathy and inability of chatbots to understand more emotional issues, especially in mental health. The responses given by chatbots were seen as depersonalized, cold, and inhuman. They were perceived as inferior to physician consultation, although anonymity could facilitate the disclosure of more intimate or uncomfortable aspects to do with health

Data Privacy
In all, 58% (15/26) of studies described the respondents' lack of trust regarding how their personal data will be collected (eg, unknowingly through voice-activated devices) and handled (eg, by whom and how) [22,[24][25][26]28,30,31,35,36,38,40,41,43,46]. However, 4% (1/26) of the studies reported no concerns regarding data sharing. This could be because of the respondents being chronic obstructive pulmonary disease patients who may have been used to their data being shared for clinical decision-making purposes [31]. Potential mitigation strategies suggested were to guarantee anonymity [26] and increase transparency in how the collected data will be used (eg, by which third party and how) [24,37]. There was a good proportion of representation from the general public, including patients (53.3%) [22,[24][25][26]30,31,37,38,46] and health care providers and IT staff (Multimedia Appendix 3).

Technology
Of the 26 studies, 6 (23%) studies discussed the participants' lack of trust in the maturity of AI technology in providing reliable and accurate information to support health-related predictions and recommendations [24,26,35,38,40,46]. This could be related to concerns over the lack of integration and synthesis of information from various sources, standardization of data collection, and the overall sustainability of AI-assisted health care service delivery [40,45]. However, 8% (2/26) of studies reported that respondents had similar trust in AI as compared with a human physician's diagnoses [28,45]. Possible mitigation strategies include increasing system transparency and reporting system accuracies [26,46]. Only 8% (2/26) of studies represented the voices of health care and IT staff [35,40,49] (Multimedia Appendix 3).

Potential Impacts of Full Automation
In all, 46% (12/26) of studies discussed the perceptions of respondents on the possibility and impacts of full automation on the health care industry, especially in terms of diagnoses, all of which reported that it is unlikely that AI will completely replace health care professionals [22,27,29,30,33,35,36,39,[42][43][44]46]. This could largely be because of the immaturity of AI technology and its limitations in providing human-like interactions (which build trust) [27]. Instead, many patients preferred a combination of both AI and human physicians in diagnoses to achieve a more accurate and comprehensive evaluation [30,39]. Most of the responses represented the voices of health care and IT staff (58.3%) [27,29,35,36,[42][43][44] (Multimedia Appendix 3).

Needs to Improve Adoption of AI in Health Care
Besides the needs highlighted to mitigate the concerns, several additional features were found to potentially improve the adoption of AI in health care (Table 3).

Enhance Personalization and Customizability
Of the 26 studies, 6 (23%) studies discussed the need for AI to personalize information such as the explanation of diagnoses, recommendations, patient education, and even pertinent questions or issues to raise to their physicians [23,24,[30][31][32]46]. Some studies also mentioned the need to customize chatbot features according to user preferences (for fixed options or free-texts) [23,24].

Enhance Empathy and Personification of AI-Enabled Chatbots and Avatars
In all, 27% (7/26) of studies highlighted the respondents' concern over the lack of empathy, which is a crucial element of human interaction to build trust between service providers and consumers. However, empathy must be displayed tactfully in verbal and nonverbal expressions such that it does not appear to be "creepy and weird," especially in populations with mental health issues [24]. Personification was also emphasized to increase the relatability, connection, and appeal to interact with the chatbot or avatar [23]. Perceived anonymity in interacting with the chatbot was also highlighted to assist in communication regarding sensitive topics [26].

Enhance User Experience, Design, and Interconnectedness With Other Devices
Overall, 15% (4/26) of studies described the need to improve user experience to increase user engagement with AI [23,25,28,31]. Strategies include needs-based interaction timing, the use of suitable verbal and nonverbal expressions, interconnectedness with other information sources (eg, electronic health record), apps (eg, calendar), and devices (eg, smart home technology-enabled devices).

Educate the Public on AI Capabilities
Of the 26 studies, 6 (23%) studies highlighted the lack of public and clinical awareness on the capabilities of AI in health care, of which the majority of the respondents expressed their willingness to learn [26,29,33,35,38,40]. A better understanding of the advantages and disadvantages of AI in health care could enhance the health care service delivery efficiency while balancing the expectations from it.

Principal Findings
On the basis of the 26 articles included in this scoping review, we identified the perceptions and needs of various populations in the use of AI for general, primary, and community health care; chronic diseases self-management; self-diagnosis; mental health; and diagnostic procedures. However, the use of AI in health care remains challenged by the common perceptions, concerns, and unmet needs of various stakeholders such as patients, health care professionals, governmental or legal regulatory bodies, software developers, and industrial providers. Simply introducing AI into health care systems without understanding the needs of stakeholders will not lead to a sustainable change [50].
Our results showed that, similar to most ITs, AI was generally favored for its on-demand availability, ease of use, potential to improve efficiency, and reduce the cost of health care service delivery. These features could enhance patients' compliance to health care treatments and recommendations that may be inaccessible or inconvenient. For example, patients are traditionally required to commit to a physician's consultative appointment that could be relatively inflexible because of a long list of patients, and one could be forced to skip the consultation because of a conflict in their schedule. AI confers the benefit of information collection and dissemination beyond the constraints of time and place, which have been shown to improve medication adherence through an AI-based smartphone app [51] and diet and exercise adherence through an AI-based virtual health assistant [52]. Our findings also demonstrated that AI is valued for its potential to speed up health care processes such as diagnosis, waiting time, communication with care teams, decisional support, and other routine tasks (eg, progress monitoring) that can be automated. This increase in service delivery efficiency frees up time and resources for clinicians to focus on tasks that involve more unexpected variabilities such as dealing with rare disease management and interacting with patients, thereby reducing the risk of burnout, job dissatisfaction, and manpower shortage [53].
Although our findings showed high rates of acceptability, concerns were raised about the lack of trust (in data privacy, patient safety, and technology maturity) and the impacts of AI-driven automation on health care job security and health care services. Ethical controversies surrounding the use of AI in health care have been long-standing. Although there are increasingly more regulatory guidelines available, such as those developed by the World Health Organization [54] and the European Union [55], the use of AI in health care remains debatable because of the challenges in ensuring data privacy and proper data use [56]. This is especially true when data collection modes are conducted through third-party apps, such as Facebook Messenger (Meta Platforms), of which privacy policies are governed by technology companies and not health care institutions [24]. Moreover, although there are privacy and security precautionary measures, the increasing reports of data leaks and vulnerabilities in electronic medical record databases erode population trust. Future security and transparency measures could consider the use of blockchain technology, and privacy laws should be properly delineated and transparent [57].
This review also found the need to enhance the personalization and customizability of information provided by AI, the incorporation of empathy and personification in AI-based conversational agents, the user experience through better design and interconnectedness with other devices and systems, and the need to educate the public on AI capabilities. Concerning personalized health care, reports generated by AI should be integrated and explained in accordance with each individual's demographic and clinical profile to facilitate self-management [46]. We also identified the need for AI to not only assist in the understanding of patients' medical condition but also the provision of relevant treatment options and personalized recommendations with intuitive actions provided (eg, a button to call an ambulance when deemed necessary by the AI) [31]. This coincides with existing studies that highlight the predictive power of AI in providing support to preventive disease onset or deterioration through interventions tailored according to user preferences [58]. For example, AI has been used to provide just-in-time adaptive interventions that prompt users to perform healthy behavior changes (eg, healthy diet and exercise and smoking cessation) based on constant data collection of their behaviors and preferences [49]. However, the data collection of users' behavioral or clinical information should also consider the customizability of input options (eg, providing predefined options or allowing for free-text input) to enhance the usability and adoption of such systems, depending on user preferences [24]. Personification of AI-based conversational agents to express human-like identity, personality, empathy, and emotions was also highlighted as an area of improvement to enhance human-chatbot interactions and eventually user adoption [59]. It was also important for the AI systems to be accessible through various devices (eg, tablets, televisions, laptops, and smart home appliances) and modes (eg, text and speech) for the convenience of information consumption and data collection. Finally, our findings suggest a need to address the knowledge deficit in the definition, capacity, and functions of AI. This could be done by cultivating AI literacy and exposure from childhood [60] and incorporating the AI curriculum in health care training and upgrading courses [61].
Overall, our study findings are consistent with well-established theories such as the Technology Acceptance Model, of which the second version proposed by Venkatesh and Davis [62] posits that technology acceptance is strongly associated with the perceived usefulness and perceived ease of use, which are influenced by subjective norms, images, job relevance, output quality, result demonstrability, experience, and voluntariness [63]. Therefore, to enhance the acceptability of AI in health care applications, its perceived usefulness over and above the current standard practices such as capacity to increase service delivery efficiency and community-based self-diagnostic accuracy should be emphasized. Such messages should be designed to be relevant to the individual and organizational adopters of a social system through various communication channels and change agents (ie, gatekeepers and opinion leaders). Such messages should be persuasive to spark five stages of adoption, namely, knowledge, persuasion, decision, implementation, and confirmation, known as the diffusion of innovation theory by Rogers [64]. Different strategies are also needed to correspond with the different categories of adopters, namely, the innovators, early adopters, early majority, late majority, and laggards. Different rates of technology adoption are associated with one's risk tolerance related to higher social economic status, education level, and financial stability [65]. An example is the case of AI adoption in chronic disease early detection and management in the United Arab Emirates. Success was attributed to the managerial, organizational, operational, and IT infrastructure factors that contribute to the factors of the Technology Acceptance Model [66]. However, advanced technologies such as AI continue to be relatively expensive and require eHealth literacy, which may widen the digital divide, and therefore the data divide and health disparity among societies. According to a report published in The Lancet, the internet remains inaccessible to approximately 50% of the global population because of a digital divide [67]. In addition, there are specific guidelines on the implementation of AI in health care service delivery, such as the quality of data and certification of AI systems, which may deter adoption [68].

Limitations
This study had several limitations. First, only articles written in English were retrieved, possibly limiting the comprehensiveness of our findings. However, we conducted a search on Google Scholar to supplement the electronic database search for more relevant papers. Second, the studies were largely heterogeneous in their study designs, research aims, and data collection methods. Third, there were limited studies on the perceptions of AI and clinical researchers who could provide outlooks on the perceptions of the general public. Finally, the public's perceptions of AI in health care may be limited by their knowledge of the definitions and capabilities of AI, as highlighted in our findings that there is a need to enhance the public's knowledge on AI. Therefore, the priority or importance of each perception and need could not be evaluated. The inclusion of articles based on our definition of AI could also have limited the scope of this study. Studies that considered different definitions of AI may have been excluded.

Recommendations for Future Design and Research
This study highlighted the perceptions and needs of AI to enhance its adoption in health care. However, one major challenge lies in the extent to which AI is tailored according to each individual's unique preference, and if such preferences are largely varied, how data can be aggregated for analyses and applicability in specific health care applications. Therefore, future studies that use AI should not only consider the issues raised in this study but also clarify the applicability in their applications and target population. A prior needs-based analysis is recommended before the development of AI systems.

Conclusions
Although AI is valued for its 24/7 availability in health care service delivery, ease of use, and capacity to improve health care service provision efficiency, concerns over trust in data privacy, information credibility, and technological maturity remain. Although several mitigation strategies such as enhancing transparency over predictive accuracy and information sources were identified, other areas of improvement were also highlighted. Future studies and AI development should consider the points raised in this study to enhance the adoption and enhancement of AI to improve health care service delivery.