McCormack and McCance [38] coined PCPF. Person-centered care is a care philosophy with people living with dementia [39]. Despite cognitive impairment, person-centered care recognizes that a person living with dementia still has personhood; should be seen as a person; and has diverse needs, such as psychosocial needs, which need to be met by care to achieve the person’s holistic well-being. PCPF provides a framework for understanding the factors influencing the practice of person-centered care with people living with dementia [38,40]. Our review focuses on adopting AI-enabled robots to care for people with dementia. PCPF is, therefore, a good fit for our review. PCPF initially focuses on nursing practice with people living with dementia [38]. However, gradually, it has been adopted by other disciplines, such as social work [41] and rehabilitation [39]. PCPF is an evolving framework since its first publication in 2001; its author continues to enhance it over the years by absorbing lessons from new research and practice [36,38,40]. We have adopted the latest version of PCPF as published by the authors in 2023 [36].

PCPF comprises 5 domains: prerequisite, practice environment, person-centered process, outcome, and macrocontext. Figure 1 presents the framework from the authors’ recent publication in 2023 [36].

PCPF considers that person-centered care practice is shaped not only by factors at the level of the patient and the health care provider but also by factors at the level of organization and society. These domains are interrelated with each other, as elaborated by the authors [36]: “to reach the centre of the framework, one must first take account of the macrocontext, followed by consideration of the attributes of staff, as a prerequisite to managing the practice environment, and in order to engage effectively through the person-centred processes. This ordering ultimately leads to the achievement of the outcome.” Under each of the 5 domains, there are constructs. There is no hierarchy among the constructs; constructs in the same domain can be related or even overlap [36].

The prerequisite domain examines the “attributes” of health care providers required to provide person-centered care, and the constructs under this domain are “being professionally competent, having developed interpersonal skills, being committed to the job, being able to demonstrate clarity of beliefs and values, and knowing self” [36]. The practice environment domain looks into “the context,” that is, the care setting and organization, where person-centered care is provided, and under this domain, the constructs include “appropriate skill mix; systems that facilitate shared decision making; the sharing of power; effective staff relationships; organisational systems that are supportive; potential for innovation and risk taking; and the physical environment” [36]. The person-centered process domain explores health care providers’ “ways of engaging that are necessary to create connections between persons (living with dementia),” and the constructs under this domain include “working with the person’s beliefs and values; engaging authentically; being sympathetically present; sharing decision making; and working holistically” [36]. The outcome domain looks into the “result of effective person-centred practice,” and the authors suggest that effective person-centered practice should “enable human flourishing for those who give care and for those who receive care” [36]. Under this domain, the constructs are that “decision-making is shared, relationships are collaborative, leadership is transformational, and innovative practices are supported” [36]. The macrocontext domain refers to regional, national, and international “factors that are strategic and political in nature that influence the development of person-centred cultures” [36]. The constructs under this domain include “policy frameworks, strategic frameworks, workforce developments, and strategic leadership” [36].

Damschroder et al [42] developed the CFIR by consolidating constructs from 19 implementation frameworks or theories in 2009. The authors updated it by incorporating new literature and feedback in 2022 [43,44]. Our review will refer to the latest version of CFIR, which is freely accessible on the CFIR website [45]. CFIR is a framework for understanding the contextual factors that influence implementation in a clinical setting or organization [44]. Our review concerns the implementation of AI-enabled robots in LTC. Thus, CFIR and our review are a good match. In addition, previous literature suggested that CFIR can be used with other frameworks [44]. In our review, we are using it with PCPF.

According to the authors, one way to use CFIR is to better understand the findings on implementation [44]. We will use CFIR to understand the findings from the literature on health care providers’ perspectives on the barriers to implementing AI-enabled robots and the strategies for overcoming them. The authors suggested 2 approaches to using CFIR to understand the findings on implementation: deductive and inductive. Our study will use a mixed approach, which will be further elaborated.

CFIR comprises 5 domains: innovation, inner setting, outer setting, individuals, and implementation process; there are constructs under each domain, and CFIR has 48 constructs in total [37]. The innovation domain explores “the ‘thing’ being implemented,” such as research, programs, policies, and innovations [37]. The constructs under this domain include “innovation source, innovation evidence-base, innovation relative advantage, innovation adaptability, innovation trialability, innovation complexity, innovation design, and innovation cost” [37]. The outer setting domain examines “the setting in which the inner setting exists,” such as the health and community organizations [37]. Under this domain, the constructs are “critical incidents, local attitudes, local conditions, partnerships and connections, policies and laws, financing, and external pressure” [37]. The inner setting domain looks into “the setting in which the innovation is implemented,” such as an LTC home or a hospital [37]. The constructs under this domain are “structural characteristics, relational connections, communications, culture, tension for change, compatibility, relative priority, incentive systems, mission alignment, available resources, and access to knowledge and information” [37]. The individuals domain explores “the roles and characteristics of individuals” involved in implementation [37]. Under this domain, the constructs include “high-level leaders, midlevel leaders, opinion leaders, implementation facilitators, implementation leads, implementation team members, other implementation support, innovation deliverers, and innovation recipients” [37]. The implementation process domain examines “the activities and strategies used to implement the innovation” [37]. The constructs under this domain are “teaming, assessing needs, assessing context, planning, tailoring strategies, engaging, doing, reflecting and evaluating, and adapting” [37].

One of our primary research questions is to explore the barriers that health care providers perceive in adopting AI-enabled robots in LTC homes. After reviewing and analyzing the existing literature, 3 barriers related to this regard were identified (Textbox 3).

Health care providers perceived the technical complexity and limitations of robots as a barrier to adopting robots. For example, in the study by Huisman and Kort [67], using the humanoid robot NAO to entertain residents and stimulate them to do physical exercises, health care providers expressed frustration with the complex robot’s operation steps and the robot’s short battery life. In the study by Hebesberger et al [68], the autonomous robot SCITOS was used to perform safety checks around the LTC home (eg, checking if doors were closed and fire extinguishers were in place) and greet visitors at the LTC home’s lobby. Nevertheless, health care providers reported that the robot’s slow response and rigid system discouraged their use of it. In the study by Pfadenhauer and Dukat [69] on the social robot PARO, health care providers mentioned that as the robot could not move independently, they had to carry it around the LTC home for residents’ use, which was inconvenient.

In addition, a lack of knowledge and skills contributes to the reluctance to adopt these technologies. For example, in the study by Papadopoulos et al [70], using the humanoid robot Pepper to interact with residents socially, health care providers reported not being fully equipped to operate and maintain the robot. In the study by Melka et al [71], using the humanoid robot NAO for rehabilitation and recreational assistance, health care providers hesitated to use the robot as they feared making mistakes because of a lack of knowledge and skills using it. Even if there is training, the willingness to participate is low due to existing heavy workloads. For instance, referring to the study by Li et al [72], using humanoid robots Pepper and NAO to entertain residents, health care providers expressed concerns about the time needed to join training as they already had heavy workloads.

Health care providers have expressed concerns, including potential negative impacts, doubts about their usefulness, and ethical implications.

Some health care providers are concerned that AI-enabled robots might negatively impact residents. For example, one of the robots used in the ethnographic study by Chang and Šabanović [73] was the robotic vacuum Roomba, and health care providers expressed concerns about residents’ risk of falling. In the study by McGinn et al [49], using the social robot Stevie to communicate with residents, health care providers were worried about hygiene hazards, such as the robot not being adequately sterilized after use and the potential spread of germs among residents. The study by Obayashi et al [74] used the social robot Sota; The health care providers were concerned that the robot’s flashing eyes might scare residents.

Some health care providers were concerned that the AI-enabled robots might negatively impact them. For example, in the study by Hung et al [11], on the use of robots, including social robot PARO, in LTC, health care providers expressed their concerns that using these robots increased their workloads, such as teaching and assisting residents to use the robots, cleaning and charging the robots, and handling technical glitches. The study by Mitzner et al [75] used service robot PR2 to assist in caregiving tasks such as medication dispensing and transferring residents. Health care providers were concerned that the robot would replace human functions. In the study by Melkas et al [71], using the humanoid robot NAO to assist in rehabilitation and recreation, health care providers were reluctant to use it as they had established workflows, and including the robot would interrupt how they had been doing things. Erebak and Turgut [76] studied health care providers’ attitudes toward robots, such as the autonomous robot Artificial Intelligence Lightweight Android (AILA). They found that the health care providers did not trust robots to be fully autonomous in decision-making. Instead, they preferred robots that allowed them to make certain decisions as humans (although the authors did not specify what these decisions would be).

Some health care providers have raised doubts about the usefulness of AI-enabled robots for residents in LTC homes. Some providers doubted whether residents would be interested in robots. For example, in the study by Robinson et al [77], on social robotic pet PARO as a companion with residents, health care providers raised doubts that residents would be interested in it because residents preferred real pets, and PARO looked like a toy. In the study by Louie et al [78], using the social robot Tangy, health care providers said residents preferred humans, and Tangy’s appearance and voice were too mechanical. In the study by Huisman and Kort [67], using the humanoid robot NAO, health care providers said that residents would be bored by the lack of choices of programs that the robots could offer for recreation and physical exercise.

Other health care providers doubted the practicality of AI-enabled robots for residents. For instance, in the study by Louie et al [78], the primary function of the social robot Tangy was to speak to residents and interact with them. However, Tangy could not speak the languages of some residents. Some health care providers pointed out that these residents could not understand and interact with Tangy. The study by Bäck et al [79] used humanoid robot NAO to demonstrate physical exercise to residents. However, health care providers hesitated to use the robot because its size was too small for residents with eyesight impairment to see it, its voice was too soft for residents with hearing impairment to hear its instructions, and residents with cognitive impairment could not follow its demonstration.

Furthermore, providers doubted that the robots would fit the LTC home context. For instance, one robot used in the study by Chang and Šabanović [73] was the autonomous mobile robot, TUG, which was designed to help health care providers deliver care and medical supplies and clean the LTC home. However, health care providers did not find it helpful because its size was too large to navigate the LTC home’s narrow hallways.

Health care providers have expressed ethical concerns about using AI-enabled robots in LTC settings. For example, in the study by Moyle et al [80], social robot PARO was used to interact with residents. Health care providers raised concerns that the use might infantilize residents due to the robot’s toy-like appearance. In the study by Lehmann et al [81], PARO was used as a robotic pet for companionship with residents. Health care providers expressed concerns that the robot could deceive residents with cognitive impairment as a real pet.

Privacy concerns have also been raised, particularly concerning surveillance. For instance, Christoforou et al [82] looked into different types of nursing and social and physical assistive robots. Some health care providers expressed the feeling that the robots were monitoring their work. In the study by Papadopoulos et al [70], using the humanoid robot Pepper to communicate with residents, health care providers worried about residents’ privacy, especially since Pepper had cameras on its forehead.

The last barrier expressed by health care providers is that LTC homes lack the resources to use AI-enabled robots. The types of resources that LTC homes lack include time, robots, infrastructure, and maintenance costs. First, as suggested, health care providers expressed concerns about the lack of time to learn, maintain, and assist residents using the robots [11]. Second, some health care providers said that their LTC homes did not have sufficient robots. In the study by Hung et al [11], a nurse recalled how 2 residents fought with each other for 1 social robot PARO. Third, some health care providers mentioned that their LTC homes lacked infrastructure for robot use. Melkas et al [71] used a humanoid robot NAO for diverse purposes (eg, rehabilitation and recreation), and NAO needed Wi-Fi. However, some health care providers mentioned that their LTC home had a poor internet connection. They also added that their LTC home did not have sufficient physical space to store the robot. Finally, keeping a robot is expensive, and some health care providers were worried about the maintenance costs. In the study by Casey et al [83], using the social robot MARIO to stimulate residents’ cognition and memories, such as giving them updates on news, health care providers were concerned about the costs of keeping the robot and suggested that it would be a better idea to spend money hiring more health care providers than keeping the robot.

To overcome the barriers, another primary question of our review is to identify the strategies suggested in the literature, including (1) accommodate the various needs of residents and health care providers, (2) increase the understanding of the benefits of using robots, (3) review and overcome the safety issues, and (4) boost interests in the use of robots and provide training. The strategies can be summarized in an acronym, “AI-ROBOT” (Textbox 4).

One strategy is to collect feedback from health care providers to design AI-enabled robots that better accommodate the various needs of residents and health care providers. For example, one of the robots used in the study by Yuan et al [84] was the humanoid robot NAO for communication with residents. Health care providers raised a concern that some residents could not understand English. In response to this feedback, Yuan et al [84] suggested incorporating songs in languages other than English into the robot to meet these residents’ language needs better. The study by Bäck et al [79] used NAO for demonstration of physical exercise. In response to the feedback from health care providers that residents with sight impairment could hardly see the robot and the residents with hearing impairment could not hear the instructions by the robot clearly, Bäck et al [79] recommended painting the arms of the robot in sharp color and giving it a loud and clear voice to accommodate the visual and hearing needs. In the study by Cavenett et al [85] who used the social robot PARO for social interaction with residents, corresponding to health care providers’ concerns that using the robot would add to their workload and interrupt their established workflow, Cavenett et al [85] proposed acknowledging the concerns and discussing with health care providers to understand their work needs and explore how the use of robot could address these needs and fit with their existing workflow.

Another strategy is to increase health care providers’ understanding of the benefits of using robots. The authors of the literature reviewed mentioned that when health care providers better understand the benefits of using robots, they will accept and use them more. For example, in the study by Huisman and Kort [67], some health care providers supported using the humanoid robot NAO because they found that it could enhance the emotional well-being of residents and bring them joy. In the study by Kolstad et al [86], some health care providers welcomed the social robot PARO because they found that it could stimulate residents’ functions, such as interactions with people. In the study by Follmann et al [87], the social robot Temi was used to contact residents’ relatives. Health care providers welcomed the robot because they found different benefits to it: it was easy for residents to use and health care providers did not need to transport the robots between residents, did not need to stand by to provide supervision when residents were using the robot, and did not need to disinfect the robot.

In addition, the reviewed literature suggests letting health care providers learn that the relationship between health care providers and robots is not competitive but complementary, emphasizing that humans perform certain tasks better than robots. A collaborative approach leverages the strengths of robots and humans, ensuring a higher quality of care. For example, Cavenett et al [85] used social robots to communicate with residents. Corresponding to health care providers’ worry that robots would replace their communication role, the authors suggested that robots only complemented health care providers’ communication role because robots were not as capable as humans to catch the nonverbal cues of residents.

The third strategy is to review the safety issues with staff and address their concerns. The study by Shin et al [88] used the robot SmartWalker to guide residents to walk. The health care providers asked if the robot was safe, considering that the residents were prone to falling. Shin et al [88] recommended incorporating safety designs into the robot. For example, when there were obstacles and stairs in front of residents, the robot would give audio warnings to the residents. In the study by Hung et al [11], in response to health care providers’ concerns about residents’ safety that residents might fight over the robots, Hung et al [11] proposed having risk assessment and management guidelines in place to avoid conflicts and violence over robots and guide health care providers on what to do in case these happen.

It is crucial to boost interests in using the AI-enabled robots from both the health care providers and the residents. The authors of the reviewed literature suggested increasing health care providers’ interest in using robots. For instance, in the study by Chang and Šabanović [89], the researchers used the social robot PARO with residents in public areas of LTC so that health care providers could see the process of using the robot and witness how the robot provided therapeutic effects to residents. The researchers found that this raised the health care providers’ interest in using the robot because they told the researchers that seeing how the researchers used the robot with the residents stimulated them to think about how they could use it in their work.

Since health care providers expressed concerns that residents might not be interested in using the robot, the authors of the reviewed literature also suggested increasing residents’ interest in using robots to address their concerns. Robinson et al [77], who used the social robot PARO, proposed changing the robot’s color from white to a more appealing color. Louie [90] recommended dressing up the social robot Tangy so that it looked more attractive. In the study by Hebesberger et al [91], using the robot SCITOS to accompany residents doing walking exercises, the researchers suggested giving the robot a name so that residents felt the robot was more personalized and thus more interested in it. In addition, in response to health care providers’ comments that SCITOS’s voice was too mechanical, they proposed giving the robot a more attractive, more natural voice.

The last strategy is to provide training. The authors of the reviewed literature recommended training and support to health care providers. Huisman and Kort [67], who used the humanoid robot NAO, proposed that the training should give health care providers clear instructions on how to use the robot for physical and recreational activities with residents. They added that time needed to be reserved for training. Otherwise, health care providers could not find time to do the training within their busy work schedules. They also recommended setting up a help desk, which health care providers could contact by phone or email when they encountered any challenges using the robot. When health care providers were more familiar with using the robot, they recommended peer learning (ie, encouraging the health care providers to support each other in using the robot). Yuan et al [84], who used the social robot PARO and humanoid robot NAO for interaction and communication with residents, raised the need for an instruction manual in place so that health care providers could refer to it after training.

As suggested, we compared our findings with PCPF and CFIR to further our understanding of the findings. The authors of PCPF [40] and CFIR [99] suggested that discussing all constructs in 1 paper is not feasible, so they recommended that the users of their framework select a few most relevant constructs to the research. We selected constructs most relevant to our review through team discussions.

In our review, one barrier to adopting AI-enabled robots is that health care providers feel that they lack the knowledge and skills to use the robots [70]. One strategy is providing training to health care providers so that they can improve their knowledge and skills [67]. The knowledge and skills needed to use robots to provide care may be part of “professional competence,” a construct under the “prerequisite” domain of PCPF. This construct refers to “the knowledge, skills and attitudes of the person to negotiate care options and effectively provide holistic care” [36].

Another barrier mentioned in our review is that health care providers are concerned that robots might replace human functions [75]. One strategy is to let health care providers know that they are better at providing these human functions, such as communication, than robots, as they are humans [75]. The constructs “engaging authentically” and “being sympathetically present” under the “person-centered process” domains of PCPF help us further understand why health care providers are better than robots. “Engaging authentically” refers to “the connectedness between people, determined by knowledge of the person, clarity of beliefs and values, knowledge of self and professional expertise” [22]. Robots cannot engage with residents “authentically” because robots are not authentic humans, despite being equipped with AI. “Being sympathetically present” means “an engagement that recognizes the uniqueness and value of the patient by appropriately responding to cues that maximize coping resources through the recognition of important agendas in the person’s life” [36]. The study by Cavenett et al [85] in our review mentioned that humans are better than robots at responding to nonverbal “cues” of residents, which are a crucial element to show residents “being sympathetically present.”

The “shared decision-making” construct under the “person-centered process” domain of PCPF refers to “engaging persons in decision-making by considering values, experiences, concerns and future aspirations” [36]. This construct made us wonder why health care providers did not mention too much about how the lack of residents’ involvement in decision-making was a barrier to the adoption of robots or suggest involving residents’ voices in the adoption process, especially since many barriers that the health care providers mentioned were related to the residents, such as potential negative impacts of robots to residents [49,73,74], perceived lack of usefulness of robots to residents [67,77,78], and ethical concerns related to residents [70,80,81]. One explanation might be that health care providers thought that they knew the residents well and could represent the voices of residents, so involving residents in decision-making was not a concern from their perspective. Future studies might examine this further. Another construct is “health and social care/policy” under the “macrocontext” domain of PCPF, which refers to “the decisions, plans, and actions that are undertaken to achieve specific health and social care goals within a society” [36]. Health care providers in our review did not mention too much about the influence of health and social policy on their adoption of AI-enabled robots, although there were international and national policies about the use of AI, such as the guidelines by the World Health Organization [100] and Health Canada [101].

One barrier to adopting the AI-enabled robots mentioned in our review is health care providers’ perceived technical complexity [68]. CFIR has a construct of “innovation complexity” under the “innovation” domain, which resonated with our findings. This construct states that “the innovation is complicated, which may be reflected by its scope and/or the nature and number of connections and steps” [37].

Another barrier is a lack of resources, such as time for health care providers to learn, maintain, and assist residents in using robots [11]; infrastructure such as Wi-Fi and storage place [71]; and maintenance costs [83]. CFIR has the construct “available resources” under the “inner setting” domain [37]. “Available resources” means “resources (that) are available to implement and deliver the innovation” [37]. This construct helps us think that the lack of resources is an organizational-level barrier, as the “inner setting” in our review context means the LTC home (ie, the organization). In other words, the organization needs to be involved in resolving these barriers.

The construct “need” is under the “individuals” domain of CFIR [37]. The “need” construct refers to the following: “the individual(s) has deficits related to survival, well-being, or personal fulfillment, which will be addressed by implementation and/or delivery of the innovation” [37]. Our findings include barriers such as health care providers’ doubted usefulness [79] and ethical concerns [81] about using robots for residents. The “need” construct helps us realize that these barriers are related to health care providers’ concerns about the “needs” of residents. For example, in the study by Bäck et al [79], health care providers doubted the usefulness of the humanoid robot NAO in facilitating exercises because its size was too small for residents with visual impairment to see it, and its voice was too soft for residents with hearing impairment to hear it. It does not meet the visual and hearing “needs” of residents. In the study by Papadopoulos et al [70] on the humanoid robot Pepper, health care providers were worried about residents’ privacy as the robot had a camera on its forehead: they were concerned about the privacy “need” of residents.

In our review, some health care providers hesitated to use robots because they were concerned that doing so would interrupt their existing workflow [71]. One strategy identified was to discuss with health care providers how to integrate the robot into their workflow [85]. The construct “compatibility” under the “inner setting” domain of CFIR helps us to understand that these are “compatibility” concerns [37]. “Compatibility” refers to the following: “the innovation fits with workflows, systems, and processes” [37].