Health recommender systems (HRS), information retrieval systems that often use artificial intelligence, are an increasingly popular method to provide patient-centric personalized health services [1-3]. HRS predict the relevance of recommendations (eg, for exercise, healthy foods, and mental health exercises) for a given user profile [4]. HRS base recommendations on what the user might like, typically using various sources of data such as personal (eg, age, gender, and socioeconomic status [SES]), health (eg, medical history and health surveys), contextual (eg, time, weather, and location), and interaction data (eg, information searched, liked, or rated) [1,5]. HRS can minimize the burden of delivering recommendations by only doing so when there is a real benefit (eg, when an individual is motivated and capable of following a recommendation) and can target multiple related behaviors at once (eg, diet, physical activity, and mental health). HRS have shown growth since 2012, particularly in the fields of lifestyle, such as personalized healthy nutrition recommendations, and personalized health promotion, such as recommending treatments of clinical diagnoses [2]. With increasing interest in using artificial intelligence to improve health promotion and prevention, and more than 50% (1 in 2 European Union citizens, aged 16-74 years) of individuals using the internet to find health information [6], the use of HRS, although still an emerging area, is expected to grow in digital health [7,8].

However, digital health interventions, including HRS, are often not developed to target health inequities, “systematic differences in the health of groups and communities occupying unequal positions in society that are avoidable and unjust” [9]. For instance, HRS have scarcely been tested in populations that are marginalized, such as those belonging to low SES and racial and ethnically minoritized individuals [10]. This is a widespread issue, including in high-resource settings, with evidence from European countries and the United States showing that White individuals and those with higher education levels, income, and English proficiency tend to have higher access to and use digital health technologies more often, instead of the groups who would benefit from these solutions most [11]. There is a lack of attention to issues of equity and inclusion, and a tendency to develop from a technology-centered (instead of a human-centered) perspective in the HRS field [2].

One way of working toward a perspective with an equity lens is the use of the socioecological model—a conceptual framework depicting spheres of influence over human behavior. This model is widely used in public health for scaling health promotion and implementation initiatives, to enable more effective public health interventions for complex social issues [12]. It has also been adapted for digital health equity [10,13]. It provides representations of how multiple, nested levels of influence (eg, community, institutional, and policy factors) interact in dynamic ways to shape individual health [14]. The socioecological model is based on a systems view of health—individual health is impacted by interactions with immediate peers and surroundings as well as broader institutional and societal norms [15]. It helps illustrate how HRS could address not just individual health needs but also the underlying conditions—such as access to health care, social support, and healthy lifestyles—that shape health outcomes. As far as we are aware, this framework has not been applied to HRS. Although some of these arguments can be applied to all digital health interventions, HRS have a unique role in the field because of their data-driven personalization and capacity to be strongly personalized due to profile- and context-crucial information.

Here we adapt the model to examine the unique potential of HRS to target and influence all its levels and thereby have great potential to promote health equity across society. We explain and illustrate how HRS can contribute to health equity by collecting data and targeting health at the systems level, integrating the context of individuals, families or peers, communities, and societies. We provide selected examples of previous work that could contribute to different layers of the framework and discuss challenges. Based on our analysis, we provide future research opportunities for HRS to target multiple levels of influence.

The novelty of this perspective is the intersection of recommender systems and public health, in particular health equity. HRS are still developing but starting to become integrated into medicine and public health [2]. There have been discussions about ethical considerations of recommender systems [16], but their potential on healthy equity, remains underdiscussed. With growing societal health inequities, and the World Health Organization’s recent call for equitable participation in a digital world [17], investing in equitable development of this new technology is essential now. As the field moves forward, we argue that an explicit health equity lens is needed from the start, not as an afterthought.

The individual level includes individual personal characteristics, such as age, education, income, and health history [19]. Using individual-level data from phones and wearables, HRS can provide personalized suggestions that integrate and recognize structural barriers to good health that marginalized groups face more often. For example, for lower SES individuals, HRS for nutrition can combine budgetary restrictions with culture and familiarity motives (eg, food associated with cultural traditions, or a sense of safety and comfort), which are more important to individuals with fewer resources and diverse cultural identities [20]. For individuals living in areas where outdoor activities may be unsafe or impossible because of weather conditions, physical activity HRS could offer alternative options, such as indoor exercises, community-centered physical activities, or free web-based classes. Furthermore, HRS can also empower users through prioritizing user decisions over system performance, which is related to digital self-efficacy, and facilitates trust in HRS. For example, in a previous study, an HRS for health coaching allowed participants to pick their own health domain to work on, and only use the information they wanted to provide, such as not using sensors or answering all health-related questions [21]. Thus, the use of HRS can lead to more equitable health care recommendations and empowered health decisions surrounding different lifestyle factors (eg, physical, mental, and social).

The interpersonal level includes an individual’s closest social circle, such as friends, partners, family members, and colleagues, all of whom can influence their health behavior [19]. HRS has the potential to positively influence the health of families, households, work environments, and other small social networks by leveraging the data of interpersonal relationships [22]. For example, HRS can provide recommendations based on what worked for others with similar health [23]. HealthNet, a patient social network platform, used a recommender system to allow patients to find others with similar conditions, learn about what treatments were useful for them, and suggest health facilities or doctors consulted by patients with similar health status [24]. However, the demographic background of patients, and the incorporation of this information into the HRS, was not reported. Basing recommendations on others with similar socioeconomic and ethnic or racial minoritized backgrounds may make HRS even more effective. Furthermore, social media network data can be used, as shown by Oliva-Felipe and colleagues [25], who developed an HRS with educational content tailored to their specific profile of dementia caregivers. HRS could also leverage peer endorsements, allowing users to see recommendations or reviews from their social connections or members of a community with similar health goals [22]. Tapping into the power of social influence and community may specifically benefit lower SES individuals by making recommendations more relevant to them and trustworthy.

The community level includes the settings in which people work, live, and have social relationships, such as schools, workplaces, and neighborhoods, which affect health [19]. HRS can play a pivotal role in facilitating health promotion and preventive strategies tailored to the community-specific risk and protective factors by capturing data specific to a community (ie, localized data collection). For example, Wayman and Madhvanath [26] describe an HRS that makes dietary recommendations based on grocery receipt data. Leveraging this kind of data, HRS developers could collaborate with local businesses, such as grocery and convenience stores, to tailor the amount of fresh fruits and vegetables guided by food purchasing data [27]. HRS developers can also work with local governments to develop systems that allow low-income residents to receive recommendations on where to shop for healthy and affordable foods. The data collected from HRS can also be used to work with the city or county to identify the most suitable walking trails, parks, and indoor exercise sites and publicize these to the community. Thus, the ability of HRS to leverage localized data provides ample opportunities for community-level interventions.

The societal level includes societal factors that influence health, such as the economic, educational, and social policies that help to create, maintain, or lessen health inequities [19].

HRS can contribute to the societal level by collecting data from individuals to build inclusive data sets with trends in health behaviors on a population level, which can be leveraged for policy making. Currently, individuals of lower SES and ethnic or racial marginalized individuals are systematically underrepresented in (open) health data sets (eg, the UK Biobank, with half a million UK participants) [28]. If HRS become widely used, by collecting data on the health of these marginalized populations in combination with contextual data, HRS can help to solve the problem of “health data poverty”, the inability of individuals, groups, or populations to benefit from discovery or innovation due to a scarcity of adequately representative data [29]. Policymakers can, for example, use these data to better pinpoint inadequate health care facilities, nutritional deficiencies, or insufficient infrastructure for healthy living and use this to devise social policies.

Regarding major public health issues like climate change, which disproportionately affects the health of marginalized populations [30], HRS could provide insight into human behavior that leads to high carbon emissions, such as food waste, or actions that can promote sustainable food consumption [31]. HRS could thus help policymakers allocate resources more efficiently by identifying areas or populations in need of specific public health interventions. This application of HRS aligns with the “Health in All Policies” (HiAP) approach promoted by the World Health Organization, in which policymakers in diverse sectors (eg, environment, agriculture, and urban planning) align efforts and health [32], consider the health impacts and benefits of plans to address employment, education, or housing. Furthermore, after implementing health policies, recommender systems could track their impact on health outcomes and provide policymakers with real-time feedback on the effectiveness of their policies [33]. Data insights collected through HRS could form the basis for more targeted policy interventions to reduce health inequities.

As we have described above in this section, HRS have immense potential to decrease health inequities. However, they are currently not designed from a health equity lens. Below we describe gaps in the current literature based on our analysis. For HRS to be more useful and accessible to marginalized populations, we need new design guidelines. Future research priorities are further described in this study that span several levels of the socioecological framework (also illustrated in Table 1).

Despite the potential of incorporating many data sources, as we explained above, most HRS do not integrate factors associated with the social determinants of health [34] and primarily use basic demographic information like age and gender, as well as individual health behaviors for tailoring recommendations [2]. As also mentioned by Richardson et al [10], in their digital health equity framework, this oversight could exacerbate health inequities, as marginalized individuals are less likely to benefit from interventions focused on behavioral factors, because of limited resources and competing priorities, which are greater in these populations. Future research should explore the integration of the social determinants of health into HRS, such as SES, healthy food scarcity, and digital and health literacy.

Because of their capability to incorporate diverse data sources, HRS are very suitable to be designed to contextualize, recognize, and incorporate diverse perspectives. Scientific knowledge in the digital health field is however currently predominantly produced through the lens of Western scholars, which marginalizes diverse and Indigenous health perspectives. In the case of HRS, users themselves have been scarcely involved in the design process. For example, in a recent review of 51 HRS, only 10 recruited users in the development of HRS testing, and the majority did not define the target group of the HRS intervention [2]. Furthermore, the majority of HRS studies are conducted in the United States and China, and low-income countries or low-resource settings are underrepresented [2].

Because of the underrepresentation of marginalized groups and lack of geographical diversity, HRS likely have cultural biases, which can make them less relevant to underrepresented groups. For example, mental health support integrated with religious or spiritual practices may be more accessible and relevant for some underrepresented groups, but it contrasts the body-mind separation of mental health dominant in a Western biomedical paradigm [35].

Decolonization involves critically re-evaluating interventions to remove cultural biases [36]. For example, mental health support integrated with religious or spiritual practices may be more accessible and relevant for some underrepresented groups, but it contrasts the body-mind separation of mental health dominant in a Western biomedical paradigm [35]. It involves centering intervention designs on the lived experience of the potential users. Thus users, especially those from marginalized backgrounds, should be included in the design and testing of HRS. This can be done through participatory research methods, where community members collaborate on all aspects of the project with researchers and other stakeholders, including feature design, testing, and dissemination [36]. This also involves examining power relationships that may underline the use of their technologies and structural and cultural factors that may broadly influence well-being [36]. Besides user involvement, technological collaborations with traditional health practitioners, such as herbalists or local healers [35] community, or faith-based organizations could help to understand and mitigate these power imbalances [37].

HRS researchers and developers should critically evaluate and report how their solutions may impact health inequities (eg, what disadvantages or privileges the research may create) [38,39]. For example, successful implementation and deployment of HRS (and many other digital health interventions) at the community level depends on structural factors such as consistent and high-quality internet access, which may be particularly challenging for resource-constrained areas [10]. If data is not available in sufficient detail or quantities, this will impact the effectiveness of HRS, leading to their lower effectiveness for marginalized populations.

Furthermore, beyond the design phase, funding for the planning and delivery of HRS to marginalized populations is necessary, to avoid that successful HRS projects fail to scale up to maturity to impact health [40]. In addition, we also need the data to track the reach and effectiveness of HRS in decreasing health inequities (metrics beyond health outcomes) [41], such as access to and quality of care, and advocate for policy changes based on these data [42]. Addressing these structural challenges and advocating for inclusive research are essential steps toward ensuring the equitable deployment and effectiveness of HRS in marginalized communities.