All quantitative analyses were performed using Stata 15. Descriptive statistics were used to summarize the characteristics of the participants and their responses to each of the three key outcomes for each of the six exemplars and in both the asymptomatic and symptomatic contexts, plus the impact of a low- or high-risk result on perspectives toward risk stratification. For analysis of the acceptability of incorporating each exemplar within risk-stratified screening, one variable was created based on the mean responses to the questions about screening frequency and starting age. Differences between asymptomatic and symptomatic scenarios were assessed using the Wilcoxon signed-rank non-parametric test to identify the significance of asymptomatic/symptomatic context on views within the sample.

Multivariable logistic regression was used to explore differences between the six exemplars within asymptomatic and symptomatic contexts and the impact of binary individual level characteristics: age (older vs. younger than 40 years), sex (female vs. male), smoking status (ever smoker vs. never smoker), ethnicity (ethnic minority vs. white), education level (university level vs. below university level), socioeconomic status decile (higher vs. lower socioeconomic status i.e. deciles 4–10 vs. 1–3), prior history of cancer (yes vs. no), technology use (at least one of using an app, tracking device or wearable daily vs. no use or use of only an electronic device daily), and perceived risk of cancer (likely vs. neither likely nor unlikely or unlikely). In all these analyses the outcomes were dichotomised by combining ‘extremely' with ‘somewhat'. The order in which exemplars were presented and an interaction between the exemplar and the order were also included.

Differences between the six exemplars are presented as predicted probabilities ±95% confidence interval using the -margins- command. The impact of individual characteristics on likelihood of uptake are presented as odds ratios ±95% confidence interval.

The free-text comments were analyzed thematically following a previously used approach (24) using Microsoft Excel spreadsheets. The comments were analyzed in 12 groups according to the dichotomy described above, and for each of the six exemplars of innovations. Comments given in asymptomatic and symptomatic scenarios were combined.

First, any comments that were generic to cancer, screening, symptoms, and/or risk assessments (that is, not specific to the innovation) or were inconsistent with the participant's previous answer (such as if they answered “very unlikely” to take part but provided a free text answer suggesting they would take up the test) were excluded.

For each group, at least two researchers familiarized themselves with a subset of the comments (20% of each group) and suggested a list of potential codes. The researchers met to agree initial code lists then piloted these on a subset of the comments (20%). After comparing and reviewing the coding decisions, they developed consolidated code lists for each group, repeating the codes across groups where possible. The remaining free text comments were then coded into the consolidated lists. Each comment was coded by two researchers and any discrepancies were discussed to reach an agreement. If a comment contained multiple meanings, it was assigned to as many codes as appropriate to cover the content.

Individual codes were then mapped into overarching themes based on the Theoretical Framework of Acceptability (TFA) (25). Four of the TFA domains were relevant—affective attitude, burden, ethicality and perceived effectiveness. A fifth domain, anticipated benefits and costs, combined the opportunity costs and anticipated benefits domains. No comments related to intervention coherence or self-efficacy. Summaries (definitions) of each code were developed and the proportion of included comments relating to each TFA domain within each exemplar was presented.

Recordings of the interviews were transcribed verbatim. For the interviews completed in multiple languages, only the parts in English were transcribed (i.e., the participant's words that were spoken through the interpreter). Data analysis was conducted using codebook thematic analysis (26), using NVivo 12 software. An initial deductive coding frame was developed based on the six exemplars of risk-based innovations and the structure of the survey, and further codes were added inductively based on participants' responses. An initial sample of the transcripts was coded by two researchers to develop and refine this coding frame. The remaining transcripts were coded by one researcher. After summarizing and reviewing each code, we defined themes according to key and repeating concepts. These were guided by the TFA where applicable (25).

An explanation of how each TFA domain contributed to participants' reasoning, with illustrative quotations, and the proportion of participants citing each reason are presented in Supplementary Table 5, Figure 2, respectively. Notably, several of the concepts applied both positively and negatively. For example, some participants were likely to take up PRS, minimally invasive tests and continuous monitoring of biomarkers because they considered them trustworthy and reliable whereas other participants would not want to take them up because they considered them not trustworthy or reliable enough. Similarly, some participants would take up wearable devices and continuous monitoring of biomarkers because they required low effort (describing them as easy, non-invasive, and convenient) whereas other participants thought they were intrusive and required too much effort (describing them as burdensome and invasive).

Overall, using innovations to assess cancer risk prior to a healthcare decision was viewed as a “proactive” approach that made “absolute sense”, with individuals generally welcoming their use. Also, the participants were positive toward each of the specific exemplars presented. However, concerns were expressed that the ease of some of the methods may lead some people to undertake them without considering the potential consequences or implications. This particularly related to minimally invasive tests, but also applied to other exemplars (Table 4; Quote 1).

The perceived level of effort required to participate was a key consideration for participants. A key benefit of AI and geodemographic segmentation was that data retrieval does not require high levels of involvement by individuals. Wearing a smart device or sensor to continuously monitor biomarkers, or providing a sample for a PRS were also viewed as low effort because they only require one appointment or setting up a device that then works autonomously. The perceived burden of minimally invasive tests was dependent on two factors: the type of test required and the test location. Urine and stool samples made some participants feel uncomfortable so they preferred saliva and blood tests. Tests that could be carried out at home were preferred to those taking place in a clinical setting. They were considered to be “less stigmatizing” and a “very natural step” (Table 4; Quote 2).

This domain tended not to be mentioned within the interviews, with the exception of concerns about the impact of geodemographic segmentation on people of lower socioeconomic status. In this context, firstly, participants were uncomfortable about the potential of causing individuals from this group to feel judged: that geodemographic segmentation may make people feel “guarded” and inadvertently cause negative views toward the decisions that people had made such as lifestyle choices, occupation, etc. Secondly, prior to correction by the interviewer, participants were concerned that geodemographic segmentation may disadvantage the poorest in society because they may not have the opportunity to move to areas that have fewer health risks. This idea was viewed as unethical and went against participants' value systems. Overall, members of the public placed importance on fair cancer healthcare, particularly for people who may already face disadvantages (Table 4; Quotes 3 and 4).

Overall, understanding was high across all risk assessment methods. Participants had prior knowledge of some of the exemplars (such as wearable devices and minimally invasive tests). Other methods were newly introduced during the interviews but were considered to be easily understood (such as continuous monitoring of biomarker levels and PRS).

In terms of coherence of the risk estimate, participants did not want to receive a risk categorization that could not be explained and expected to receive, at a minimum, a general explanation as to why they received their risk score. A lack of explanation would make them “untrust[ing]” of the NHS, particularly if they had a high risk of cancer. In particular, they would feel more positive toward the use of AI if a report could explain and justify their result (Table 4; Quote 5). This principle also applied more generally: by understanding the components of their risk score, they would treat it in a more serious manner, which could motivate people to engage in behavior change to lower their cancer risk or “make important decisions about their health” in conversation with healthcare professionals.

Furthermore, communication strategies were viewed as important in terms of the acceptability of cancer risk assessments to the public. Participants suggested that any changes to current practice should be communicated through public campaigns, letters formally stamped with the NHS logo, and conversations with GPs as the point of contact. More specifically, conflicting opinions were expressed regarding the communication of the results from individuals' risk assessments. Numeric scores were viewed as precise but had the potential to be “morbid” and it could lead to comparison amongst friends and family. Presenting the risk score categorically (high, medium or low risk) was equally conflicting as although it may be easier for the public to understand, individuals on the lower or higher ends of these scales could get either a false sense of security or anxiety.

A variety of opportunity costs were associated with different exemplars of risk assessments. For example, participants considered the implications for health insurance associated with geodemographic segmentation. Specifically, they worried if they were found to be at higher risk of cancer based on where they lived, it could cause companies to raise insurance premiums that may, in turn, exacerbate inequalities (Table 4; Quote 6). Additionally, it felt inevitable that continuous monitoring would be accompanied by an invasion of their privacy. This was particularly true for wearable or mobile technologies that would show lifestyle factors. However, they felt that the benefits outweighed invasion of privacy (Table 4; Quote 7).

Notably, participants considered the impact of AI on their relationship with healthcare professionals (the “human vs. technology” trade-off) with mixed emotions. Positively, some participants felt that AI would be more accurate than a GP. It could also be more efficient, which, in turn, may lead to positive financial implications and reduced workload and time pressure. In contrast, other participants expressed a strong preference for maintaining the traditional GP-patient relationship, valuing the interactions and conversations with healthcare professionals. They felt that the AI algorithm might generate a less accurate risk assessment because GPs could notice additional, external factors that an algorithm may miss. Participants with concerns about AI therefore did not want their healthcare to be informed solely by the output of the algorithm. They felt that AI should be used as a side-by-side or secondary check to ensure other risk assessments did not miss anything and to confirm the risk categorization (Table 4; Quotes 8 and 9).

Irrespective of exemplar, participants spontaneously mentioned concerns for data access and storage. Health data is sensitive information, therefore, they were accepting of it being accessed as long as it was for a risk assessment, as intended, and that it was safeguarded appropriately to prevent misuse (Table 4; Quote 10). Participants generally trusted the NHS with their medical data, and felt it was logical to use it to assess risk of other health conditions, not just cancer. They also expressed an interest in having access to their own personal data.

Participants were quick to share their perspectives on the effectiveness of the exemplars. Perceived effectiveness was impacted by the type and frequency of data collection, in addition to anticipation of errors and inaccuracies to be introduced within data-based innovations.

Participants tended to place a greater value on the collection of biological data than lifestyle or environmental data. It was seen to be “more sophisticated” and objective because it examines individualized internal processes and therefore also incorporates the effects of lifestyle and environmental factors. Nevertheless, both were valued: geodemographic segmentation alone could be too generalized or generic (Table 4; Quote 11) yet PRS would not take important environmental and lifestyle factors into consideration; for example, an individual may have a low genetic risk of developing cancer but live in a polluted environment or lead an unhealthy lifestyle, which could place them at higher risk (Table 4; Quote 12). Furthermore, the potential of biological data extended beyond cancer prediction, such as alerting individuals to biological indicators of other diseases, as well as benefiting society by providing further information on disease processes.

When considering the frequency of data collection, continuous methods were favored over one-off snapshots. Continuous data collection was viewed as comprehensive, reliable and accurate as it provides real time data (Table 4; Quote 13). Although single data collection events required less effort, upon reflection, participants felt that they had more potential for inaccuracies as the results reflect a person's health at just one moment in time (Table 4; Quote 14). They were also concerned that additional, non-fixed factors may impact the results. For example, the quality of the test conducted, potential for contamination, or additional biological and lifestyle related factors that change regularly (sleep, stress, diet etc). On the other hand, concerns about the accuracy and trustworthiness of wearable devices were heightened by the need to remove it and disrupt constant wear (e.g., to charge it or to wash).

For AI, participants were concerned that medical records might not be up to date. This applied to individuals not attending their GP regularly or only attending if they became symptomatic, and in cases where health events had not been reported and therefore would be unknown to the algorithm (Table 4; Quote 15). Participants proposed that a standardized check of health records should be conducted prior to running AI to ensure that there is a consistent baseline level of information and that no biases are present within the algorithm. They also worried that individuals may not report geodemographic segmentation data correctly leading to errors in risk categorization (Table 4; Quote 16).

Self-efficacy was not considered for AI, continuous monitoring of biomarkers, geodemographic segmentation or PRS. These risk assessments tend not to require an individual to perform a behavior independently for data collection.

Conversely, participants showed a lack of confidence that they or others could complete the tests that required more active engagement. As described for the burden of completing certain minimally invasive tests, some were unsure about providing urine or stool samples at home. Moreover, they worried that a lack of self-efficacy would impact the accuracy of risk assessments based on data collected by wearable devices, which could lead to an inaccurate risk category (Table 4; Quote 17). These technologies might be difficult for individuals who are not “tech savvy”. Participants felt that they may not be able to correctly operate the technologies or understand the outputs, which could lead to them feeling “overwhelmed” and “agitated”.

Lastly, throughout the interviews it became apparent that the level of receptiveness to risk assessment methods varied according to individual tendencies, irrespective of the exemplar risk assessment method they reviewed. Some participants demonstrated an avoidant coping style toward health and therefore were not enthusiastic to take up a risk assessment as they “would rather not know”. In contrast, some participants proactively “want[ed] to keep on top of their health” and therefore were enthusiastic about taking up the innovations and finding out the results. Others explained that such decisions should be taken in conjunction with other family members (Table 4; Quote 18).

We observed moderate or high overall intended uptake and acceptability of the use of risk-based innovations to stratify care in both asymptomatic and symptomatic contexts; this is consistent with previous studies that have shown public willingness to participate in cancer risk assessments (19, 27, 28). Analyses of interview and free text data highlighted underlying positive attitudes toward such approaches: they are considered logical, intuitive and likely to be effective. Of note, though, acceptability of risk stratification tended to be slightly lower than likelihood of risk assessment uptake. This suggests that participants were more eager to learn of their cancer risk than for it to be used to inform screening or investigations for symptoms (more so in the symptomatic context). Although we were not able to explore why this might be in the qualitative analyses, it is known that the public anticipate gaining knowledge of their personal risk of cancer to be empowering and motivational for behavior change (19, 29, 30), as well as benefitting health by enabling those at high risk to access increased monitoring and/or prophylactic intervention (31–34). Risk assessment in the symptomatic context may additionally be considered a form of investigation and more participants considered the effort required to complete a risk assessment worthwhile when symptomatic.

Differences in views according to individual participant characteristics are important to consider for implementation of risk-based innovations, alongside individual differences that are not easy to gauge such as avoidant or proactive tendencies. Across the three main outcomes, participants were statistically more likely to respond positively if they reported use of health apps. This was particularly true for wearable devices. Other associations with participant characteristics varied by outcome and context. For example, female participants were less likely to find risk-stratified healthcare acceptable, but there were no differences compared to males for anticipated uptake. Conversely, participants of minority ethnicity and low socioeconomic status were less likely to say they would take up a risk assessment for screening but did not find risk stratification significantly less acceptable than their comparators. This is consistent with socioeconomic and ethnic inequalities in uptake of screening itself (35–37). Similar patterns have been found for returning a fecal immunochemical test kit in symptomatic patients (38), help-seeking for possible cancer symptoms (39), and in a study assessing uptake of breast cancer risk assessments based on phenotypic and genetic risk and mammographic density (40). It is therefore necessary to ensure that those already less likely to engage in screening and present with symptoms are not discouraged further through the prospect of a risk assessment.

Almost all participants were likely to take up the offer of minimally invasive tests, find risk stratification based on the results acceptable and be comfortable with the results being kept on record (a minimum 85.0% across all outcomes), making them the most popular risk assessment method examined. Qualitative findings suggest that this was driven by understanding that such tests would require low effort, particularly where they could be completed in comfort at home rather than needing an appointment, and were perceived as effective, reliable and accurate due to their biological nature. This aligns with some of the reasons recent focus group participants gave for their enthusiasm for multi-cancer early detection blood tests: their familiarity with and the relatively non-invasive and standardized nature of blood tests (41). Likewise, discrete choice experiments have reported minimally invasive tests (e.g., stool tests, breath analysis and blood tests) to be preferred over more invasive tests for population cancer screening, providing that sensitivity is sufficiently high (42, 43). As such, while our study is the first to ask participants to consider such tests for cancer risk assessments rather than screening itself, the findings are consistent with views on their use in other contexts.

The second innovation based on biomarkers—continuous monitoring of biomarkers—was regarded similarly positively to minimally invasive tests, albeit with perceptions of a greater burden yet higher potential effectiveness. This may explain why 12.8% more participants were likely to take up the offer in the symptomatic rather than asymptomatic scenario. There is a lack of information about perceptions of continuous biomarker monitoring in the context of cancer healthcare. Observational research is needed to understand to what extent experiences in other contexts such as continuous glucose monitoring in people with diabetes, where empowerment and reassurance are experienced alongside burdens due to device complexity and the non-stop nature (44, 45), would be experienced if it were used for cancer risk assessments. Nevertheless, our findings suggest that the public would be open to this idea.

The use of wearables in healthcare is a rapidly growing area of interest to both individuals and policy makers, both in those with cancer diagnoses and healthy or asymptomatic individuals (16). There is a paucity of data on acceptability in certain areas relating to cancer; for example, while there is promise in the feasibility, efficacy and acceptability of wearable devices to promote physical activity in cancer survivors, public perspectives toward portable, non-invasive technologies for early breast cancer detection remain unknown as research has focused on technological developments so far (13, 46). In this study we found that participants were moderately receptive to the use of wearable devices compared to the other exemplars included. They described positive aspects including the acceptable burden because data are collected automatically once the device is set up, which may only take one appointment as for some of the other innovations, and/or they may already own a wearable device (e.g., a smartwatch). This aligns with the observed association between receptiveness and familiarity with health apps. Participants also expressed a range of potential added benefits from motivating behavior change to altruistically contributing to medical knowledge. Ethicality due to the focus on lifestyle factors was the greatest concern; specifically, the potential to attribute blame for a high cancer risk on lifestyle choices as well as privacy concerns over continuous surveillance. This issue is of paramount importance to policymakers and researchers too (47).

Participants were receptive to PRS with 84.7% and 88.1% participants likely to take up this test in asymptomatic and symptomatic contexts, respectively, with more certainty in uptake in the symptomatic context. This is comparable to findings on the acceptability of PRS for risk assessment in hypothetical contexts (48–51), as well as 84 to 96% uptake observed in research settings (52). Our study reveals that the low burden of PRS was the key driver in this context: participants were encouraged by the requirement to only provide one sample at home or at a single appointment. Furthermore, only a few participants reported concerns about PRS, although these are important to consider as they have been observed in other settings. Some raised issues of ethicality or data protection (52). Others considered a potential cost in which access to other tests may be delayed or withdrawn whilst waiting for the PRS result or upon receiving a low genetic risk of cancer. Although fear of delays could be mitigated by appropriate policies, the concern about being “penalized for good genes” echoes public objections about risk stratification for those with lower cancer risk and again highlights the value of screening within society (19). Lastly, other participants expressed mistrust of the accuracy of the algorithms, methodology or technology.

Receptiveness to AI analysis of medical records fell in the midpoint between the other exemplars. The qualitative work suggests there were mixed views accounting for the different perceived impacts of AI on relationships with healthcare professionals: some considered AI favorably, suggesting that it could outperform or help their overworked GP, whereas others were concerned that it was more fallible, would negatively affect this valued relationship or would have unintended impacts such as deskilling GPs. This echoes public perspectives toward AI in a multitude of clinical contexts, albeit not specifically for analyses of medical records, where authors of a systematic review report that “the acceptability of AI generally hinged on its use as a support, rather than a replacement, of healthcare providers” (53). It was also important to our participants that they could understand how their risk classification came about, which put some people off the prospect of AI. A recent review suggested that inability to understand an AI algorithm reduced patient autonomy in decision making and led to greater psychological harms caused by misdiagnoses (54). Nevertheless, amongst our participants, the minimal burden meant there was little to lose. Mixed perspectives as reported here were also observed in a recent survey of 600 members of the US public: 52.0% were comfortable with AI being used to predict future medical conditions compared to, for example, 84.2% who were comfortable with its use in administrative tasks such as scheduling appointments (55). Together, these findings suggest that using AI to estimate risk of cancer based on general practice medical records has the potential to be acceptable to the public provided that suitable safeguards are in place.

Finally, we observed that geodemographic segmentation was the least preferred option with only approximately two-thirds of participants considering that they would be likely to take it up and fewer finding it acceptable to inform healthcare, especially in the symptomatic context. Interestingly, this was the example in which the order effect was greatest: if participants saw it first, they were more receptive to it. This suggests geodemographic segmentation was considered as a somewhat good idea, but not as good as the other exemplars. As the least preferred option, implementation efforts may be best focussed on other innovations while further research is conducted into the utility of geodemographic segmentation and public perspectives toward it.

Reasons for likely uptake included the low burden. Although the impact of environmental factors on cancer risk resonated with many [as in other research where pollution and radiation were perceived to have some of the highest impacts on cancer risk (56)], this was not an obvious part of demographic segmentation to others. The interviews revealed misunderstanding in which participants associated personal finances with socioeconomics. They therefore assumed that personal finances would inform the cancer risk assessment and healthcare, which raised issues of ethicality and negative affective attitudes. This misunderstanding could be corrected by the interviewer, but not in the survey that was completed independently. As such, the population-level acceptability of geodemographic segmentation should be interpreted with this in mind. It also highlights two important implications: it is necessary for policies to be coherent and fair for society in order to be acceptable. For example, the message that screening and referral informed by geodemography may help to reduce inequalities could be received positively, subject to evaluation (57).