Uppsala municipality includes 22 neighborhoods, with around 52,000 inhabitants, identified as socioeconomically disadvantaged [26]. These neighborhoods have higher unemployment rates, a higher proportion of residents receiving financial assistance, higher illness rates, lower household income, a higher proportion of economically vulnerable children and youth, lower percentage of residents with a postsecondary education [26], and a high proportion of non-European immigrants [27]. The robot screening was set up in one of these 22 neighborhoods, while the study participants came from 5 of the 22 neighborhoods. The robot screening was therefore conducted in the setting and population intended for its future implementation. To the best of the authors’ knowledge, this study is the first of its kind to explore users’ perception of the CMD risk assessment with a social robot in a community context.

This was an exploratory qualitative study where data were collected through semistructured interviews [28]. The SRQR (Standards for Reporting Qualitative Research) checklist is reported in Checklist 1. The study is guided by a pragmatist research paradigm. Pragmatism focuses on what works in practice and views knowledge as context-dependent and shaped by experience [29]. This fits the study well, as the goal was not to test a theory or discover universal phenomena but to understand how participants experienced their interaction with the robot in a specific context. The choice of framework analysis with mixed deductive-inductive coding reflects this stance: UTAUT was used as a practical tool to structure the analysis, while inductive coding allowed for new themes to emerge directly from the data [30].

The study was conducted as part of formative studies in the PREVENT project, which aims to reduce the risk for CMDs in high-risk populations within socioeconomically disadvantaged neighborhoods in the municipality of Uppsala through cocreating, implementing, and testing community-based support for healthy and active living [31].

The social robot used for this study was the “Furhat” robot, developed by a Stockholm-based company, consisting of a head (which can display different facial features and expressions) and a loudspeaker [32]. The expressions and movements thereby mimic human behavior. It can be programmed to have fluid conversations in different languages [32].

Participants were selected using purposive sampling with specific inclusion criteria to ensure relevance to the study’s aims. Through this variation, participants’ characteristics, experiences, and perceptions could be explored and compared between members of different groups. Eligible participants were aged between 30 and 75 years, lived in a selected area in Uppsala, and spoke Swedish (easy Swedish which is often spoken by people new in Sweden). Three participants also had a translator with them who helped them during the interaction and the interview. The age range of 30-75 years was chosen to reflect the demographic most at risk for CMDs, as individuals aged 30 years and older from regions such as the Middle East, Sub-Saharan Africa, and Southeast Asia have been shown to have a higher risk of developing CMDs [33]. The upper age limit of 75 years was set to reduce the likelihood of including individuals with neurodegenerative conditions, such as dementia or cognitive impairment, which could interfere with their ability to engage with self-reporting screening tools [34]. Easy Swedish was included as a requirement to minimize the complexity of introducing a social robot with speech capability to a multicultural and low-income suburb where many different languages are spoken.

The screening and interviews were conducted over a 1-week period from October 19 to 26, 2023. Each participant was scheduled to conduct the screening and the interview during a specific time slot in this week. The risk assessment and interviews took place at the locale of a local civil society organization. For this study, the social robot “Furhat” was programmed to conduct the risk assessment using the standard Finnish Diabetes Risk Score (FINDRISC) questionnaire [35]. FINDRISC is a validated tool to assess people’s risk of developing T2D and, by extension, CMDs, within the next 10 years [36,37], due to common risk factors and causative mechanisms. It is an effective noninvasive tool used for screening in both health facilities and outside, such as community settings. FINDRISC includes a total of 8 questions that include demographic variables (age), anthropometric measures (BMI and waist circumference), behavioral measures (fruit and vegetable consumption and physical activity), and personal and family medical history (intake of antihypertensive drugs, high blood glucose, and family history of diabetes). Responses from each individual are scored, and a score of 13 or higher indicates a high risk for T2D [35] and is found to correlate closely with worsening values across various clinical and anthropometric parameters related to risk for T2D and CMDs [37]. The risk assessment with the robot was conducted in easy Swedish. Most interviews were conducted in Swedish or English, but 3 participants also had a friend or family member to help with translation to their mother tongue.

During the robot screening, only the male team member in charge of the technical aspects of the robot (“wizard”) and the participant remained in the test room. Participants were informed of this arrangement before providing consent. The test was set up in a way that the robot and the wizarding station were on opposite ends of the room to make participants focus on the robot instead of the wizard (Figure 1). During screening, the wizard tried to interfere as little as possible. However, some participants asked him when they did not understand something, in which case he responded. In one case, the robot failed to register voices, so the wizard entered the answers on his laptop manually. This was, however, not noticed by the participant as the robot was still able to ask questions. After the risk assessment, the interviewer returned to conduct the interview, while the wizard left. The interview was conducted next to the robot, so that participants could look at it during the interview to remember their feelings and experiences during the robot screening more easily (Figure 1). After the interview, the FINDRISC scores were explained to participants.

The research process from obtaining consent to robot screening and the interview that followed took approximately 45 minutes per participant. Sample size was assessed using the concept of information power. Based on 4 of its 5 items (a narrow aim, the specificity of the experience studied, application of an established theory, and quality of dialogue based on strong and clear communication on a narrow topic), we assessed the 13 interviews to hold sufficient information power for the purpose of this study [38].

The interview guide used for this study included questions about participants’ demographics, their experiences with robot screening, suggestions for improvement, and their believed usefulness of robot screening in the community (Multimedia Appendix 1). The interviews were conducted by 1 member of the research team who was also responsible for fieldwork in the study area within the PREVENT project (JIH). All interviews were audio-recorded.

JIH is a Swedish-born, male anthropologist. At the time of the interviews, he had conducted 1 month of full-time participatory fieldwork in the study setting documented in field notes [39]. The fieldwork revolved around informal “deep hanging out” [40] in participants’ everyday lives. Details on the fieldwork and fieldwork activities have been published elsewhere [41]. With the exception of 2 participants recruited through other channels, JIH had met all participants more than once before the interviews and had informal interviews with them. This prior relationship may have facilitated rapport and openness during the interviews [42]. At the same time, the familiarity between JIH and participants may also have shaped both what participants chose to share and how their accounts were interpreted during the interview. Thus, while JIH contributed ethnographically grounded insight into the data, SDB, who led the analysis of the collected interview material, brought interpretive distance from the fieldwork and interview encounters. This division of roles supported a reflexive analytic process in which interpretation could be grounded in local contexts while also being questioned from a less embedded position.

The interviews were transcribed in Swedish and translated into English. Since some of the participants had a translator with them, 3 interviews had passages in Arabic, which were transcribed in Arabic and translated to English. All transcriptions and translations were performed by trained university students, native or fluent in the corresponding languages. Interviews were transcribed verbatim, with spot checks and back translations performed on the first transcripts (particularly those in Arabic to confirm the correct meanings). The interviewer later reviewed the translated transcripts to ensure that the translations retained the meaning and content of the original interviews. The framework analysis method was used to analyze the data in accordance with the UTAUT framework [30].

The UTAUT framework consists of 4 core determinants—performance expectancy, effort expectancy, social influences, and facilitating conditions (Figure 2)—that directly affect behavioral intention and use behavior. Behavioral intention describes the user’s intention to use the technology in the near future, while the actual use of the technology is called use behavior in the framework [11]. Performance expectancy is defined as the extent to which users believe that using the technology will help them achieve gains, in this case for their health [11]. Effort expectancy describes the user’s ease of interaction with the technology [11]. Social influences refer to how much the user believes that important others think they should use the technology [11]. Facilitating conditions refer to the user’s belief that there is technical or organizational infrastructure in place to support them in using the technology [11].

For the coding, a mix of deductive and inductive approaches was used. An initial codebook was created by SDB, using a deductive approach guided by the UTAUT framework and informed by the interview guide and recurring topics identified during initial familiarization with the data [30]. The 4 UTAUT determinants served as the basis for the initial codebook. This codebook was then used by JIH, MG, and SDB to independently code 1 transcript. After discussing differences and similarities in the coding, the codebook was revised by SDB. Where participant responses did not fit within these predefined categories, new codes were added inductively, allowing themes to emerge directly from the data. The coding was conducted across 4 rounds with the help of the open-source software “QualCoder (version 3.4)” [43], after which the initial UTAUT-based themes were recategorized due to overlaps, resulting in a revised set of themes that better reflected the data. The codebook can be found in Multimedia Appendix 2.

To ensure trustworthiness, strategies addressing credibility, dependability, confirmability, and transferability were applied. Credibility was supported through investigator triangulation: 1 transcript was independently coded by 3 researchers (SDB, JIH, and MG) using an initial codebook, after which discrepancies were discussed and reconciled. This helped identify individual interpretive biases before coding the remaining data. Dependability and confirmability are accounted for by keeping an audit trail of the whole coding process, changes made in the codebook, and in the categorization of codes, as well as field notes from the interviewer and the wizard. Field notes and personal observations were used to continuously reflect on assumptions and minimize bias during the writing process. Transferability of the results was supported through a thick description of the sampling strategy and the research context [44].

The study received ethical approval from the Swedish ethical review authority (RN2022-04034-01). All participants provided informed written consent and were compensated with vouchers from supermarkets in the neighborhoods. Participants were informed about how their personal data were collected and stored. Interviews were audio-recorded and the data were anonymized. All data were stored on a secure Uppsala University server to which only the research team had access. Every person who transcribed, translated, or worked with the transcripts signed a confidentiality agreement.

A total of 13 people participated in the study. All 13 participants were born outside of Sweden and had Swedish as their second language. The participants were born in Syria, Iran, Saudi Arabia, Sudan, Somalia, Eritrea, Yemen, and Kurdistan. Participant characteristics are shown in Table 1.

Participants generally stated that they perceived the robot screening as positive but nevertheless most would have preferred doing the risk assessment with a human instead. Two major contributors for this preference were the language barrier between participants and the robot, and participants’ perception of the robot to be lacking in comparison with a human because of its limited programmed abilities, lack of emotions, and lack of responsive adjustment to people’s needs. However, participants generally perceived the robot as nonjudgmental, trustworthy, and keeping users’ information confidential, which had a positive effect on participants and made them feel more comfortable interacting with the robot for the risk assessment. Certain groups of people such as women, older adults, or immigrants were often perceived to be more reluctant or having a harder time using the robot, which was mostly not supported by the accounts of the respective groups.

Most studies investigating users’ interaction and perception of social robots or socially assistive robots are conducted with older adults living in retirement homes. The use of social robots in preventive and community-based health care settings has had very little attention so far. However, similar to this study, other studies found that users generally had a positive attitude toward using social robots but perceived them to have limited capabilities in interaction compared with humans [45,46]. Users especially criticized the robots’ inability to show emotions, display different facial expressions, respond to the users’ mood [46], or having only a limited amount of interactive functions [45], which is in line with the findings from this study. A study by Wu et al [47] in a gerontology ward in France found that participants had a negative attitude toward the socially assistive robot that was tested. They reported feeling uneasy with the technology, feeling stigmatized by others for using it, and did not perceive it as useful [47]. Although most participants in the PREVENT study reported a preference to interact with humans, they generally expressed a more positive attitude toward the robot. While some of the users of this study reported that they were afraid of others ridiculing them for using the robot or they were slightly distrusting of the robot, the majority expressed the opposite and stressed potential benefits of the robot screening in the community.

The findings of this study about the robot being perceived as friendly, trustworthy, nonjudgmental, and maintaining confidentiality are in line with findings from other studies where a social robot was tested in a retirement home [48] or as a language teacher [49]. A study by van Maris et al [50] found that users’ trust in robots increases over time. Some of the participants of the PREVENT study reported feeling slightly stressed since interacting with a robot was something new and unusual for them. Therefore, it is possible that repeated interactions with the social robot for health screening could enhance users’ trust and acceptance of it over time.

While most participants perceived women and men to be equally interested in the robot screening, women were sometimes perceived by men as preferring human interaction. The women themselves, however, refuted this view of them as being less interested in using a social robot for health screening. What is interesting here is not so much a gender difference in acceptance but rather the assumptions about emotions as motivators for seeking or avoiding risk screening with a robot. These assumptions reflect broader gendered stereotypes about comfort, relationality, and technology use. Existing literature shows mixed findings on gender preferences in social robot acceptance, with some studies reporting higher acceptance among women [51] and some among men [52,53]. Such differences may partly relate to the appearance and abilities of the robots tested, but our findings suggest that perceptions of emotional motivation may be equally influential in shaping expectations about who would use such technologies.

When it comes to age, participants in this study believed older individuals to be more hesitant toward interacting with the robot, whereas younger individuals were believed to be more interested and more competent with technology. While some of the older participants reported their age to be a hindering factor in the interaction, others believed that it was more a matter of one’s own mindset and interests. The age stereotype is also contradicted by a study by Kuo et al [53], which investigated attitudes of users toward a health care robot in New Zealand. Older people expressed similar positive attitudes toward using the robot as middle-aged participants, although older participants were less experienced with technology [53]. A study by Van Dijk [54] discovered that the motivation to use and acceptance of technology by older adults increases if the device is perceived to be useful and convenient. Therefore, the generally positive attitude of older people toward the social robot used in this study might be influenced by their belief about its usefulness and the ease of interaction with technology.

Another group that was seen as more reluctant to use the social robot for risk assessment if it was installed in the community were immigrants. However, the participants refuted this view themselves with their enthusiasm for and interest in the risk assessment with the robot. It is important to note here that all participants were volunteers, which means they were likely among the most open and technology-interested members of their community. Therefore, they may not represent those who would indeed prefer human interaction. Nevertheless, the contrast between participants’ assumptions about others and their own personal experiences suggests that stereotypes and societal norms may shape expectations more than actual user preferences. A systematic literature review on technology use among immigrants found a limited use of technologies in older immigrants compared with younger ones [55]. Older immigrants were found to be less interested in learning how to use certain technologies, as well as having more difficulties in using them [55], which is in line with the beliefs of the participants of this study about older immigrants. An aspect of the assumed disinterest among immigrants for robot-assisted risk screening was in how they included themselves or not in their own definitions or descriptions of immigrants. Participants often excluded themselves from the groups of immigrants when talking about preferences—defining “immigrants” as those who struggle with language or cultural integration. This aligns with Moffitt and Juang’s [56] research, whose participants described immigrants as people who “don’t speak the language well” or “stick only to their own cultural norms”—criteria they used to exclude themselves from the label.

The UTAUT framework, developed to explain users’ intentions to adopt a technology and their subsequent usage behavior [11], was valuable for structuring the deductive part of our analysis. It provided a useful lens to interpret how factors such as performance expectancy, effort expectancy, and social influence shaped participants’ experiences with the robot and ultimately shaped their behavioral intention and use behavior [11]. With regard to performance expectancy, participants, independent of age or gender, generally believed that using the social robot for the risk assessment will bring them benefits for their health. However, this positive perception was mostly associated with the community screening itself and not the robot screening per se. As for the effort expectancy, the language barrier was reported as the main barrier in the ease of interaction with the robot. The perception that it would be easier to communicate with a human might have contributed to participants’ preference of conducting the screening with a human. The language barrier was not dependent on participants’ gender, although older age was attributed with having more difficulties in learning Swedish. When it comes to social influences, some participants reported being scared that others would ridicule them for interacting with the robot, which could potentially make them more reluctant to use it. However, most participants felt safe and comfortable during the interaction with the robot. These feelings of safety and trust emerged as an important inductive finding, as they are not included in the UTAUT core constructs. These experiences did not differ by age or gender of participants. With regard to facilitating conditions, participants’ suggestions for improvements to the future installation of the robot—such as enhanced language support and more interactive and emotionally responsive features—highlight the infrastructural and contextual factors perceived as necessary for successful deployment. These findings illustrate how UTAUT can capture important aspects of acceptance in this setting. However, the UTAUT framework was initially developed to model behavioral intention and sustained technology adoption in contexts of repeated use, which is not applicable to a periodic activity such as cardiometabolic risk screening. The findings should therefore be understood as reflecting initial impressions rather than settled adoption attitudes, particularly for constructs such as performance expectancy and social influence, which are likely to be shaped by repeated exposure and experience over time. Participants’ beliefs about the robot’s benefits for their health or about how others in their community would react to it may shift considerably following multiple encounters, and a longitudinal study may therefore yield different findings. Furthermore, the framework did not fully explain participants’ broader beliefs about who in their community would (or would not) be willing to use the robot, which suggests that additional perspectives are needed to fully understand social meanings around robot-mediated health screening.

This study had several limitations, the main one being that the risk assessment with the social robot was conducted in Swedish, which was not the mother tongue of the participants. However, this being the first robot deployment in the community, use of Swedish as the operational language was a deliberate choice to reduce the complexity of interactions and the scope of the study. Since it was the first field deployment, we kept the interaction elements simple, with no added interpretative or explanatory elements where advanced language skills would have been necessary for both the robot and the participants. For this reason, the robot communicated the FINDRISC result as a numerical score only, without categorizing it into a corresponding risk level (eg, low, moderate, and high). While this simplified the interaction, it may have limited participants’ understanding of their diabetes risk, as the clinical significance of a given score is not necessarily self-evident without an accompanying risk category. The language bias was minimized by having a translator for the interviews, where possible. Consequently, the language barrier was a foreseeable outcome of the study design rather than an unexpected finding. Furthermore, it remains unclear whether the language difficulties reported by participants were specific to interacting with a robot in Swedish, or whether they would apply equally to any Swedish-language health encounter. Future studies deploying multilingual robot capability would help disentangle these 2 aspects.

Second, the wizard was present in the room and visible to the participants during the screening, although he placed himself outside their direct line of vision. Although he tried to interfere as little as possible, some participants instinctively turned to the wizard when they did not understand a question, which might have influenced participants’ ease of interaction with the robot. Additionally, some participants reported discomfort answering sensitive health-related questions in front of the wizard, which may have affected their responses. Furthermore, the wizard’s presence may have had broader consequences for the validity of the findings. Knowing that a human was available as a fallback could have inflated participants’ feeling of comfort and trust, as the interaction did not reflect a fully autonomous robot deployment. This reflects the early-stage nature of this deployment, where human oversight remains necessary to ensure technical reliability. Findings on acceptance and trust should therefore be understood as reflecting robot screening in a supervised context.

A further limitation concerns the conditions under which the interviews were conducted. Participants were interviewed immediately after the screening, in the same room as the robot, and by a researcher involved in the PREVENT project. These conditions may have encouraged socially desirable responses, as participants might have felt inclined to respond positively out of politeness toward the interviewer or reluctance to criticize the technology in its presence. This could partly explain the pattern observed in the findings, where participants frequently described the robot very positively while simultaneously expressing a preference for human interaction. Future studies could address this by introducing a time delay between the screening and the interview.

Furthermore, selection bias needs to be considered when interpreting the findings. Since all participants were volunteers, they were likely among the most willing and open members of their communities. This is particularly relevant for theme 1, where participants reflected on whether others in their communities would use the robot. Interestingly, some participants distanced themselves from the groups they were describing; for example, participants who were themselves immigrants often excluded themselves when talking about the assumed preferences of immigrants. While this reflects a nuanced and self-aware perspective, it also highlights that the views of those who might genuinely hesitate are not captured in this sample. The perspectives of those who declined to participate or would never have volunteered remain unknown, and claims about community-wide acceptance should therefore be interpreted with caution.

Moreover, as UTAUT was developed to model sustained technology adoption over time, findings based on a single postinteraction interview reflect initial impressions rather than considered adoption attitudes. Longitudinal follow-up studies examining whether initial perceptions translate into actual acceptance of social robots for health screening over time would strengthen the evidence base.

This study provides insights into users’ perceptions of the usability and acceptance of a social robot for risk assessment deployed in a community context. Although participants enjoyed the interaction with the robot, perceived it as nonjudgmental, and would recommend it to others in their community, most would prefer conducting the screening with a human. Both the language barrier experienced during the robot-participant interaction and their wish for more interactive, emotionally responsive, and communicative features of the robot contributed to this preference. However, this needs to be considered in the context of developing and implementing cost-effective processes for at-scale population-based screening, which would most likely necessitate the minimization of human personnel.