Many authors speculate that these technologies will create positive health outcomes for populations (Althobaiti, 2021; Cheng et al., 2020; Abramoff et al., 2021; Castagno and Khalifa, 2020; Kelly et al., 2020). Some of these positive expectations have been associated with specific optimization of various health services. Authors have mentioned the gain in terms of accessibility (Fornasier, 2019; Xie et al., 2020; Bates et al., 2018; Jones et al., 2020). The combined use of AI and BD opens a new scalability and the possibility to treat an unimaginable quantity of patients in comparison to what the actual workforce can offer (Abramoff et al., 2021). In that sense, AIS can offer a response to the actual health workers shortage that many health systems are facing. In parallel, these technologies could reduce the cost of health services (Kern et al., 2016; Grigorovich and Kontos, 2020) and make resource allocation more efficient (Stylianou and Talias, 2017; Galetsi et al., 2019; Sparrow and Hatherley, 2019; Bates et al., 2018; Machluf et al., 2017; Canaway et al., 2019; Peters and Buntrock, 2014). These benefits could be significant for low- and middle-income countries (LMICs) (Alami et al., 2020), where AIS could complement existing health services (Schwalbe and Wahl, 2020).

Authors have identified specific interventions that could be optimized with the integration of AI and BD such as helping to manage disease (Kerr et al., 2018), faster (Fornasier, 2019) and with more precision (Althobaiti, 2021), or otherwise facilitate diagnosis (Noorbakhsh-Sabet et al., 2019; Sparrow and Hatherley, 2019), determining appropriate treatments (Yang and Chen, 2018), e.g., with the use of precision medicine (Ahmed et al., 2020), and improving patient outcomes more generally (Canaway et al., 2019). Robots more specifically could help reduce loneliness (Miller, 2020) and otherwise induce positive emotions in older patients (Ienca et al., 2016), enhance their autonomy and thus reduce the burden on the healthcare system (de Graaf et al., 2015; Ienca et al., 2016). At the population level, AI and BD can support proactive interventions, particularly in populations of lower socioeconomic status (Machluf et al., 2017; Eng, 2004; Zhang et al., 2017), improve the prevention, prediction and treatment of chronic diseases (Lajonchere, 2018; Kern et al., 2016; Cool, 2016), make disease screening more efficient (Morgenstern et al., 2021), and facilitate epidemics surveillance (Galetsi et al., 2019; Cheng et al., 2020; Bates et al., 2018; Roberts, 2019) and the decision-making in cases of global health emergencies (Galetsi et al., 2019). AIS can offer more targeted populational interventions through so-called “precision public health” (Johnson, 2020; Dolley, 2018). Authors also noted benefits for healthcare systems including analyzing their inefficiencies (Manrique de Lara and Pelaez-Ballestas, 2020; Ho et al., 2020), detecting problems in health laboratories (Yang and Chen, 2018), facilitating the assessment of health technologies and drugs (Lajonchere, 2018) and streamlining the workflow (Thorpe and Gray, 2015a; Tang et al., 2018; Sparrow and Hatherley, 2019; Kostkova et al., 2016). Finally, the AI and BD technologies could optimize the research process at the very core of healthcare (Mooney and Pejaver, 2018; Sparrow and Hatherley, 2019; Balas et al., 2015; Joda et al., 2018) and facilitate the distribution of its benefits (Ahmed et al., 2020).

Conversely, authors have identified numerous negative health outcome that could be aggregated into two clusters. The first one focusses on the errors that could be introduced by AI and BD. System dysfunction or malfunction are part of the game (Satava, 2002) and an error in AIS used systemically in healthcare could lead to harming 1,000 of patients (Sparrow and Hatherley, 2019). There is a possibility of misdiagnosis because of bugs or the overreliance of healthcare professionals (HCP) on AIS (Morley et al., 1982). The efficiency of AIS can lead to lower the human scrutiny on the system and diminish human capacity to control the system (Sarbadhikari and Pradhan, 2020). Another risks is the use of an AIS for a purpose other than what it was designed for (Ahmed et al., 2020). In the same vein, the vulnerability of these systems for cyber-attacks could disrupt the use of AI devices and affect populations (Abdulkareem and Petersen, 2021). Errors do not only pertain to the systems and HCPs can misleadingly interpret the results (Stylianou and Talias, 2017; Satava, 2002), misleadingly interpret the results of AIS because of their reluctance or distrust AI predictions (Ahmed et al., 2020; Horgan and Ricciardi, 2017; Nebeker et al., 2019).

The second cluster highlights the reductionist view of health introduced by these systems and the risks that something important will be missed (Dolley, 2018). Careless use may lead to wrong results, harming populations (Alemayehu and Berger, 2016) and wasting resources (Green and Vogt, 2016). The central role of BD for AIS risk reducing populations to numbers, narrowing the whole human experience (Sparrow and Hatherley, 2019) characterized by, inter alia, its irrationality, unpredictability and vulnerability (Kerr et al., 2018; Brill et al., 2019; Ahmed et al., 2020; Delpierre and Kelly-Irving, 2018; Prainsack, 2020), as well as its cultural dimension, situatedness, and its reliance on values, preferences and beliefs (Mentis et al., 2018; Kee and Taylor-Robinson, 2020; van Deursen and Mossberger, 2018). This form of dehumanization (Althobaiti, 2021; Ienca et al., 2016) can be detrimental to the therapeutic relationship (Sparrow and Hatherley, 2019) by providing services devoid of human contact (Fulmer, 2019; Kernaghan, 2014; Miller, 2020; Cordeiro, 2021), the empathy and the compassion normally offered by HCP (Kerr et al., 2018; Morley et al., 1982; Manrique de Lara and Pelaez-Ballestas, 2020). Also, these technologies are seen as ways to ground the personalization of medicine based on the individual’s genetic background. For some, it is feared that this narrow use will divert the focus away from public health interventions, and from upstream determinants of health (Kee and Taylor-Robinson, 2020; Kenney and Mamo, 2019).

Because of the scale at which it is used (Abramoff et al., 2021), some hope that AIS used in population health interventions will contribute to reducing health disparities (Zhang et al., 2017; Genevieve et al., 2019), but the reverse effect is anticipated by many (Nebeker et al., 2019; Terrasse et al., 2019; Breen et al., 2019; Holzmeyer, 2021; Luk et al., 2021). Some fear that these technologies will affect disproportionately parts of the population (Zhang et al., 2017; Montgomery et al., 2018) such as people with disabilities (Jones et al., 2020), vulnerable populations (Rosen et al., 2020), and marginalized communities (Hu et al., 2017). This could be partly due to interventions (e.g., precision public health) narrowly focused on biomedical factors and surveillance instead of taking into consideration social determinants of health (Johnson, 2020; Mentis et al., 2018; Kenney and Mamo, 2019; Backholer et al., 2021; Trein and Wagner, 2021). Conversely, public health surveillance programs may unduly focus on vulnerable populations because they may have less control over their “digital footprint” (Rosen et al., 2020), be insufficiently prepared to represent their interests (van Deursen and Mossberger, 2018) and lack time to manage their virtual identity (Montgomery et al., 2018). In parallel, there is a risk that the technology be used with bad intentions, perpetuating social prejudices and therefore increase health disparities (Abdulkareem and Petersen, 2021). For example, discriminatory uses of BD and AIS, such as selecting who has access to healthcare (Sparrow and Hatherley, 2019), identifying noncompliant patients (Moutel et al., 2018) and cherry-picking patients (Zhang et al., 2017; Terrasse et al., 2019), could increase health disparities by depriving populations who need it most from access to health services (Cahan et al., 2019).

Another type of justice consideration regarding BD and AIS relates to discrimination and stigmatization. Data breaches; loss of privacy; public information on social media; the identification of individuals, falsely or not, with a medical condition, a particular genotype, or as the source of an infection (Raza and Luheshi, 2016; Shachar et al., 2020); and the inclusion of social determinants in electronic health records (Goodman, 2020) and tracing apps (Mbunge, 2020); all these situations raise risks of stigmatizing individuals and communities (Aiello et al., 2020; Galetsi et al., 2019; Hunt et al., 2020; Genevieve et al., 2019; Breen et al., 2019; Luk et al., 2021; Baldassarre et al., 2020; Mikal et al., 2016; Vayena et al., 2015; Yang and Chen, 2018; Celedonia et al., 2021; Ngan and Kelmenson, 2021; Straw, 2021; Xing et al., 2021) as well as risks of discrimination (Cordeiro, 2021; Salerno et al., 2017; Yeung, 2018; Gilbert et al., 2019; Chen and See, 2020; Jalal et al., 2020) by insurance companies and employers (Stylianou and Talias, 2017; Cahan et al., 2019; Docherty and Lone, 2015; Colloc, 2015; Salas-Vega et al., 2015; Benke and Benke, 2018; Sparrow and Hatherley, 2019; Brill et al., 2019; van Deursen and Mossberger, 2018; Manrique de Lara and Pelaez-Ballestas, 2020; Montgomery et al., 2018; Salerno et al., 2017; Sun et al., 2020; Babyar, 2019; Ajunwa et al., 2016; Adkins, 2017; Casanovas et al., 2017; Ienca et al., 2018; Mootz et al., 2020; Tigard, 2019; Rajam, 2020). These risks apply even to individuals who have not participated in research activities (Kim et al., 2017) (e.g., when members of a group have shared identifiers) (Manrique de Lara and Pelaez-Ballestas, 2020) and when data has been anonymized (Dolley, 2018; Sanchez M and Sarria-Santamera, 2019). Discrimination could also occur on the basis of race, sex (Abdulkareem and Petersen, 2021; Manrique de Lara and Pelaez-Ballestas, 2020), gender (Zou and Schiebinger, 2021), income, age (Abramoff et al., 2021), and it can take many forms such as “invisibility, exclusion, or complacency employed to avoid detection, critique, or questioning” (Dankwa-Mullan et al., 2021). At the clinical level, protocols based on population statistics may exclude the individual preferences of patients (Dagi, 2017).

One distribution consideration relates to the fair return of results of technology development. Authors highlight the risk of “digital colonialism” where privileged populations benefit from the development of technology while the less privileged are left apart. This issue can take many forms that are mostly illustrated by the unequal relationships between high-income countries and LMICs. One fear is that researchers from high-income countries take advantage of data collected by researchers in LMICs for their own advantage and without acknowledging the latter’s work (Ballantyne, 2019; Car et al., 2019; Howe and Elenberg, 2020). At the population level, some worry that health data be analyzed in high-income settings with no possibility for LMIC to control how it is used (Andanda, 2019) and to benefit from it (Dolley, 2018; Li and Cong, 2021). Digital colonialism can also take the form of AIS being developed with data from high-income countries that will lead to detrimental and discriminatory effects on health care in LMICs. For example, these AIS may recommend a health intervention that is not feasible locally or only available at a significative costs outside the country (Alami et al., 2020; Bhattacharya et al., 2021; Demuro et al., 2020). Another consideration is that there might be socioeconomical barriers that prevent the implementation, in a LMIC, of an algorithm created in a high-income country (Liu and Bressler, 2020). A corollary is “ethics dumping,” which is “exporting unethical research practices, for example, unethical data processing […] to countries where research ethics committee oversight is lacking” (Samuel and Derrick, 2020). Some could justify this “ethics dumping” with the fact that the access to healthcare can be difficult in some LMICs. In the same vein, there is a concern that non-compliant technologies could bypass security and privacy vulnerabilities since informal healthcare is more prevalent in LMICs countries (Alami et al., 2020). However, this could lead us to a new “medicine for the poor” in the same way that most of the medical equipment being sent to LMICs fail or do not work (Alami et al., 2020).

The “digital divide” argument offers a variation on the unfair distribution of outcomes issue (Zhang et al., 2017). It describes inequalities in access to data (van Heerden et al., 2020) and technologies (Bennett, 2019) caused either by of a lack of resources (Mbunge, 2020) or knowledge (Aiello et al., 2020; Galetsi et al., 2019; Bennett, 2019; Vollmer Dahlke and Ory, 2020; Genevieve et al., 2019; Lodders and Paterson, 2020). The increased use of BD and AI in health could worsen the digital divide (Eng, 2004) and perpetuate health inequities (Genevieve et al., 2019; Murphy et al., 2021) by leaving out people who cannot or do not want to use those technologies (Kerr et al., 2018; Cordeiro, 2021; Mootz et al., 2020; Fleming, 2021). This can particularly affect people in LMICs (Brill et al., 2019; Mentis et al., 2018; Manrique de Lara and Pelaez-Ballestas, 2020; Mbunge, 2020) but also populations with lower socioeconomic status in high-income countries (Budd et al., 2020). The digital divide could have multiple consequences. First, it could lead to unrepresentative data sets by excluding populations who have least access to technologies (Cahan et al., 2019; Docherty and Lone, 2015; Benke and Benke, 2018; Aiello et al., 2020; Vollmer Dahlke and Ory, 2020; Dolley, 2018; Delpierre and Kelly-Irving, 2018; Ossorio, 2014; Genevieve et al., 2019; Breen et al., 2019; Mikal et al., 2016; Yeung, 2018; Gilbert et al., 2019; Hodgson et al., 2020). Second, these populations have higher burdens of disease (e.g., advanced age, lower economic status, etc.) but have less resources to benefit from BD and AI innovations (Larkin and Hystad, 2017; Sun et al., 2020; Strang, 2020). Third, the digital divide could also create inequities in digital surveillance (Aiello et al., 2020) and be exacerbated by the uses of the technologies at the international level (Manrique de Lara and Pelaez-Ballestas, 2020). However, populations with lower digital literacy could also be overrepresented because they “may be more likely to unknowingly imply consent” (Mikal et al., 2016; Demuro et al., 2020). Programs aiming to curb the digital divide could create a “privacy divide” if they require that vulnerable populations trade their personal data in exchange for products and services (Montgomery et al., 2018).

An important concern relates to the presence of biases in datasets and the coding of algorithms that may lead to an unfair distribution of the benefits and burdens of the technology in the population or between populations. Biases may have different sources such as the obliteration of certain groups in the datasets used to train AI. This could come from observational, and sampling bias at the basis of data gathering (Cahan et al., 2019; Howe and Elenberg, 2020; Bhattacharya et al., 2021; Strang, 2020; Goldsmith et al., 2021; Tan et al., 2020) or missing data from less represented populations (Cahan et al., 2019; Lanier et al., 2020; Howe and Elenberg, 2020; Bhattacharya et al., 2021; Strang, 2020; Goldsmith et al., 2021; Tan et al., 2020). Biases in programming (Cahan et al., 2019; Wang and Alexander, 2020; Xie et al., 2020; Ahmed et al., 2020), for its part, may come from the amplification of previous biases and the failure to recognize them in subsequent stages (Morgenstern et al., 2021; Zou and Schiebinger, 2021; Baclic et al., 2020; Thomasian et al., 2021). They could also come from the erroneous decision to apply data from one population to another (Tang et al., 2018; Sparrow and Hatherley, 2019; Zhang et al., 2017; Morley et al., 1982; Delpierre and Kelly-Irving, 2018) or developers’ incorrect assumptions and beliefs (Terrasse et al., 2019; Yeung, 2018). All this will result in biased results, or to what authors refer to with the expression “garbage in, garbage out” (GIGO), meaning that biased data leads to biased results (Cahan et al., 2019; Howe and Elenberg, 2020; Evans et al., 2020). The risk at this stage is the perpetuation of biases, as biased algorithms could exacerbate already present racial and socioeconomic inequalities and vulnerabilities (Sarbadhikari and Pradhan, 2020; Luk et al., 2021; Couch et al., 2020). This may affect the health of individual patients (Tang et al., 2018; Morley et al., 1982; Sarbadhikari and Pradhan, 2020) and, moreover, the wellbeing of the global population (Docherty and Lone, 2015; Galetsi et al., 2019; Breen et al., 2019; Yeung, 2018) in terms of the perpetuation of discriminatory racial and social practices (Cahan et al., 2019; Kee and Taylor-Robinson, 2020; van Deursen and Mossberger, 2018; Altenburger and Ho, 2019) or health inequities (Cahan et al., 2019; Lanier et al., 2020; Sparrow and Hatherley, 2019; Zhang et al., 2017; Manrique de Lara and Pelaez-Ballestas, 2020; Yeung, 2018; Ballantyne, 2019; Villongco and Khan, 2020; Carney and Kong, 2017). Biased AIs seem unavoidable (Goodman, 2020), or hard to minimize (Gilbert et al., 2019), because of the black-box nature of many AIS (Lanier et al., 2020; Sparrow and Hatherley, 2019; Rajam, 2020), and the ubiquitous nature of AI (van Deursen and Mossberger, 2018). The mistake may be to consider data as pure and objective realities (Holzmeyer, 2021) although they are determined (like health and wellbeing) by economic, social, and political dynamics (Johnson, 2020; Delpierre and Kelly-Irving, 2018) as well as generating social consequences (Alami et al., 2020).

AIS using BD may affect positively individual behaviors by empowering patients in taking care of their own health (Fornasier, 2019; Lajonchere, 2018; Fulmer, 2019; Cordeiro, 2021; Manrique de Lara and Pelaez-Ballestas, 2020; Snell, 2019; Montgomery et al., 2018; Prosperi et al., 2018). These technologies could help individuals to monitor their own health (Wang and Alexander, 2020; Kenney and Mamo, 2019; Car et al., 2019), offer pertinent health information (Prosperi et al., 2018; Kostkova et al., 2016), contribute to decision-making (Bellazzi, 2014; van Deursen and Mossberger, 2018), assist in the management of their health and illness (Eng, 2004), and open the possibility of robotic assistance and interactions (Bennett, 2019; de Graaf et al., 2015; Vollmer Dahlke and Ory, 2020; Belk, 2020; Kernaghan, 2014). All this could be of great use for chronic disease management (Kenney and Mamo, 2019; Car et al., 2019), and supporting disabled people (Jones et al., 2020) or elderly people’s autonomy (Manzeschke et al., 2016). The autonomy offered by these systems may modify the power relationship with the HCP in favor of the patient (Galetsi et al., 2019; Sparrow and Hatherley, 2019; Terrasse et al., 2019; Gilbert et al., 2019), thus diminishing medical authority (Lupton and Jutel, 2015). This conception of empowerment and engagement is a strong dimension of the digital health rhetoric (Lupton and Jutel, 2015).

The combined use AI and BD can also have positive effects on collective behaviors. Some anticipate that these technologies offer platforms for collective engagement, for example in disease surveillance (Genevieve et al., 2019; Sun et al., 2020) and in the research process (Manrique de Lara and Pelaez-Ballestas, 2020; Manrique de Lara and Pelaez-Ballestas, 2020; Gossec et al., 2020; Anisetti et al., 2018; Katapally, 2020). In that vein, some see the possibility of citizen engagement in the development of these very same technologies (Casanovas et al., 2017; Altenburger and Ho, 2019), yet a lot has still to be done (Conrad et al., 2020; Breen et al., 2019; Gilbert et al., 2019; Pepin et al., 2020; Nichol et al., 2021; Tang et al., 2018; Manzeschke et al., 2016). This possibility raises its very own ethical issues regarding the authenticity of the engagement of citizens, patients or populations (Bennett, 2019; Vollmer Dahlke and Ory, 2020; Ballantyne, 2019; Evans et al., 2020).

Conversely, many speculate that the technologies will promote disempowerment. For some, patients may feel a loss of agency toward the decision taken by HCPs and AIS (Morley et al., 1982; van Deursen and Mossberger, 2018), particularly if the AIS is opaque (Amann et al., 2020), and create forms of nudging (van Deursen and Mossberger, 2018). Aside, the pervasiveness of the technology may discourage individuals to engage in their own health and leave this task to the technology (Kasperbauer, 2021). On the other hand, they may feel responsible for their health, creating “individuals on alert” (van Deursen and Mossberger, 2018; Samerski, 2018). Despite presenting themselves as patient-empowering, self-diagnosis apps still recommend users to seek medical advice, challenging patient empowerment in face of medical authority (Lupton and Jutel, 2015).

To sustain the empowering of populations and attain positive health outcomes, digital and ethical literacy appears to be an essential precondition for the stakeholders of population health (Benke and Benke, 2018; Fulmer, 2019; Delpierre and Kelly-Irving, 2018; Ballantyne, 2019). First, there is a need to educate the public regarding digital technologies using BD and AI (Galetsi et al., 2019; van Deursen and Mossberger, 2018) and their various pitfalls such the limitations of the technology (Grigorovich and Kontos, 2020; van Deursen and Mossberger, 2018; Lupton and Jutel, 2015), the complexity of privacy protection (Mikal et al., 2016; Lodders and Paterson, 2020), the risks of cybersecurity (van Deursen and Mossberger, 2018) and the inherent biases of the technology (Mikal et al., 2016). The same necessity for digital and ethical literacy by the general population has been said for policymakers (Lanier et al., 2020), HCP, and researchers (Stylianou and Talias, 2017; Leyens et al., 2017; Aiello et al., 2020; Galetsi et al., 2019; Brill et al., 2019; Machluf et al., 2017; Satava, 2002; Babyar, 2019; Gossec et al., 2020; Pepin et al., 2020; Hemingway et al., 2018; Godfrey et al., 2020; Ho and Caals, 2021). At last, ethical literacy may be critical for data scientists to achieve their aim (Dolley, 2018; Kern et al., 2016; Zhang et al., 2017; van Deursen and Mossberger, 2018; Goodman, 2020; Nebeker et al., 2019).

The question of data ownership asks the question of who can exercise power over the data that will be used to train and fed AI. The question of ownership is a complex one (Gilbert et al., 2019). Data are created by many people and all have some rights over the data (Mooney and Pejaver, 2018; Tang et al., 2018) while promoting different agendas (Alemayehu and Berger, 2016). BD derived technologies amplify this situation with their capacity, sometime furtive, to aggregate numerous sources of data. These sources of data may be as diverse as ordinary internet-connected object (Miller, 2020; van Deursen and Mossberger, 2018; Ajunwa et al., 2016) to public health surveillance interventions (Hodgson et al., 2020). All this may be complicated by the limited knowledge of the data individuals implicitly share (Horvitz and Mulligan, 2015).

A strong line of thought suggests that there is an information asymmetry between individual and corporation in the favor of the latter (Sun et al., 2020). Health data can be seen as a profitable investment for corporations (Canaway et al., 2019; Cheung, 2020). There is the possibility that private corporation owns sensitive health information (Tigard, 2019) and that they capture health data coming from public health interventions (Aiello et al., 2020). Although they might be regulated (Ballantyne, 2019), they might be less accountable for the use of data (Andanda, 2019) while caring less for the social good (Terrasse et al., 2019) than the protection of intellectual property (Salas-Vega et al., 2015) and developing monopolies (Ladner and Ben Abdelaziz, 2018; Roberts, 2019; Satava, 2003). This situation opens fear of abuses (Benke and Benke, 2018; Delpierre and Kelly-Irving, 2018; Prainsack, 2020) which makes some believe that the deployment of AIS will benefit the corporation rather than the populations (Adkins, 2017) and perpetuate social inequalities (Andanda, 2019).

While corporations play a central role in data economies, the control of individuals over their own data also need to be considered (Andanda, 2019; Colloc, 2015; Salas-Vega et al., 2015; Benke and Benke, 2018; Tang et al., 2018). Policies may play an important role in protecting this form of control (Montgomery et al., 2018) to respond to constant risk of reidentification (Delpierre and Kelly-Irving, 2018) and commodification (Conrad et al., 2020). Traditionally, patients have not been able to control their healthcare data (Bates et al., 2018), but, because of the strategic role data plays. There is an increasing demand from individuals to have access to their own data (Stylianou and Talias, 2017; Hodgson et al., 2020).

An alternative to previous mode of property could be find in collective ownership of data such as “data sovereignty” which could be defined as the “rights of a nation to govern the collection, ownership and use of its own data” (Ballantyne, 2019). It is argued that people using AIS should have the chance to have some control over their data (Cordeiro, 2021; Casanovas et al., 2017), particularly if we consider that the data provided for the development of BD and AIS is a community investment (Oravec, 2019). As a community investment, it may warrant financial returns or a stake in the decision-making (Oravec, 2019). Differences in data systems between countries raise challenges and opportunities for State-bodies (Machluf et al., 2017). Governmental control can be seen as a more secure (Snell, 2019) alternative to commercial management. Governance innovations include “data custodians and/or indigenous data governance bodies” (Ballantyne, 2019). This control of data by communities may contribute to guarantee the inclusion of diverse dimensions and include social determinants of health (Dankwa-Mullan et al., 2021). Although, this community control may be illusory if, at the end, data are stored in the cloud through a network of foreign servers (Colloc, 2015).

A related set of issues to the ones of ownership relates to data management (Stylianou and Talias, 2017; Salas-Vega et al., 2015; Galetsi et al., 2019; Wang and Alexander, 2020; Ladner and Ben Abdelaziz, 2018; Ahmed et al., 2020; Young, 2018). Authors ask what ethical data management would look like (Samuel and Derrick, 2020)? Data management is an important consideration because of the increasing number of people involved in data collection (Vogel et al., 2019), and the enormous amount of data generated (Bellazzi, 2014). Data management implies a long continuum from data production, storage, curation, analysis, protection and circulation. It raises the issue of who has the power to manage the data and the risk of centralized or commercial data control (Manrique de Lara and Pelaez-Ballestas, 2020; Tupasela et al., 2020). On the contrary, centralization may be replaced by over fragmentation and make it difficult to locate when data are used as part of large platforms or by many entities, e.g., in research settings (O’Doherty et al., 2016). In terms of population health, the more acute concern is to optimize their use (Cahan et al., 2019), because of their medical importance (Casanovas et al., 2017), and their role in eliminating health disparities (Carney and Kong, 2017). Authors sometimes talk about the stewardship of data (Sanchez M and Sarria-Santamera, 2019) which includes the “safeguards, audits and operational protocols” (Sanchez M and Sarria-Santamera, 2019).

One risk associated with data management are conflicts of interests (COI) that can arise if data belong to actors who have diverging interests. For example, corporations, governments, the public, healthcare systems, HCPs and researchers may all have diverging needs, interests and goals, raising risks of COI (Salas-Vega et al., 2015; van Deursen and Mossberger, 2018; Manrique de Lara and Pelaez-Ballestas, 2020; Salerno et al., 2017; Casanovas et al., 2017; Car et al., 2019; Altenburger and Ho, 2019). This situation may be more patent for regulators who want to promote, at the same time, commercial and public interests (Tupasela et al., 2020). At last, COIs can be hidden within the programming of their algorithms (Car et al., 2019).

Corollary issues regard the accessibility of the data and data sharing (Stylianou and Talias, 2017; Wang and Alexander, 2020; Benke and Benke, 2018; Galetsi et al., 2019; Ladner and Ben Abdelaziz, 2018; Ahmed et al., 2020; Nebeker et al., 2019; Goodman, 2020; Sun et al., 2020; Ienca et al., 2018; Tigard, 2019; Hodgson et al., 2020). Publicly funded data and data of public utility may have a stronger obligation for being accessible (Tang et al., 2018; Goodman, 2020; Gossec et al., 2020). Access to data is necessary in order to realize BD and AI’s potential for improving global (Li and Cong, 2021; Galetsi et al., 2019) and individual health (Zhang et al., 2017; Kostkova et al., 2016; Raza and Luheshi, 2016; Li and Cong, 2021; Mahlmann et al., 2017). The accessibility of data can be essential for public health, and become critical during infectious outbreaks and (Budd et al., 2020; Ballantyne, 2019). Data sharing is also strategic for research activities (Galetsi et al., 2019; Zhang et al., 2017). For example, easier access to publicly funded clinical datasets could help reduce data-access inequities between researchers (Zhang et al., 2017), and enable reproducible research (Gossec et al., 2020).

Because of its benefits, some people believe that individuals have a duty to share their data in order to advance health goals. Some authors defend the idea that it is a societal responsibility to act accordingly (Green and Vogt, 2016). In other words, individuals have a duty to share their information for the sake of their own treatment (Salerno et al., 2017), for epidemiological reasons (Salerno et al., 2017), for the advancement of health research (Tsai and Junod, 2018) or for the learning health system (Sparrow and Hatherley, 2019). Individuals will benefit at a certain point in life from these goods (Tsai and Junod, 2018) or they will contribute to the common good (Snell, 2019). Otherwise, it may be considered as selfishness or free riding (Snell, 2019).

However, this imperative to share data may face several barriers that may be practical (Leyens et al., 2017; Wang and Alexander, 2020; Ahmed et al., 2020; Balas et al., 2015; Lee and Yoon, 2017), cultural (Leyens et al., 2017; Tang et al., 2018; Car et al., 2019), economical (Tang et al., 2018; Lee and Yoon, 2017; Machluf et al., 2017), technical (Lajonchere, 2018; Noorbakhsh-Sabet et al., 2019; Zhang et al., 2017; Raza and Luheshi, 2016; Yang and Chen, 2018; Dagi, 2017; Hemingway et al., 2018; Salerno et al., 2017; Deshpande et al., 2019), political (Machluf et al., 2017; Car et al., 2019; Carney and Kong, 2017), ethical (Bates et al., 2018; Machluf et al., 2017; Balas et al., 2015; Yang and Chen, 2018; Gilbert et al., 2019; Prosperi et al., 2018; van Heerden et al., 2020), and regulatory (Salas-Vega et al., 2015; Lajonchere, 2018; Zhang et al., 2017; Manrique de Lara and Pelaez-Ballestas, 2020; Rosen et al., 2020; Raza and Luheshi, 2016; Yang and Chen, 2018; Casanovas et al., 2017; Mahlmann et al., 2017; Deshpande et al., 2019). Many stakeholders may have an interest in accessing data, e.g., researchers, health-policy makers, HCPs, insurances (Stylianou and Talias, 2017). It raises numerous questions. Who should be given access to the data? For which aim? In which conditions? (Stylianou and Talias, 2017; Casanovas et al., 2017) With which safeguards? (Andanda, 2019; Lajonchere, 2018) In which sustainable infrastructure? (Raza and Luheshi, 2016; Pepin et al., 2020) How should benefits and risks of data sharing should be distributed equitably? (Brill et al., 2019; Ballantyne, 2019) These questions are entangled in the web of issues at the intersection of privacy protection, control over data access, and protecting informed consent (Sanchez M and Sarria-Santamera, 2019).

Privacy issues are central to the ethics of AIS and BD because of the informational nature of these technologies. They refer mostly to wrongful uses of data (Colloc, 2015; Balas et al., 2015; Shah and Khan, 2020), accidental disclosure (Mooney and Pejaver, 2018; Balas et al., 2015; Conway, 2014; Casanovas et al., 2017), data crossing (Sparrow and Hatherley, 2019; Horvitz and Mulligan, 2015) or unintentional disclosure of sensitive information (Xing et al., 2021; Murphy et al., 2021; Gilbert et al., 2020). These are frequently analyzed through the lens of cybersecurity issues (Stylianou and Talias, 2017; Tanti, 2015; Dolley, 2018; Liyanage et al., 2014; Lajonchere, 2018; Salas-Vega et al., 2015; Wang and Alexander, 2020; Heitmueller et al., 2014; Kerr et al., 2018; Xie et al., 2020; Galetsi et al., 2019; Sparrow and Hatherley, 2019; Fulmer, 2019; Althobaiti, 2021; Bates et al., 2018; Machluf et al., 2017; Ahmed et al., 2020; Canaway et al., 2019; Ienca et al., 2016; Kostkova et al., 2016; Balas et al., 2015; Prainsack, 2020; van Deursen and Mossberger, 2018; Ossorio, 2014; Rosen et al., 2020; Montgomery et al., 2018; Salerno et al., 2017; Gilbert et al., 2019; Sun et al., 2020; Ajunwa et al., 2016; Casanovas et al., 2017; Ienca et al., 2018; Mootz et al., 2020; Tigard, 2019; Dagi, 2017; Hodgson et al., 2020; Lupton and Jutel, 2015; Pepin et al., 2020; Manzeschke et al., 2016; Belk, 2020; Snell, 2019; O’Doherty et al., 2016; Deshpande et al., 2019; van Heerden et al., 2020; Cutrona et al., 2012; Fornasier, 2019; Hoffman and Podgurski, 2013; Terry, 2014; Torous and Haim, 2018; Tsai and Junod, 2018; Veiga and Ward, 2016). Privacy breaches are increasingly observed in the health sector (Sun et al., 2020; Ajunwa et al., 2016; Dagi, 2017) and they have been highlighted at different phases of health data circulation from collecting (Mohr et al., 2017), transferring between linked services (Shah and Khan, 2020), sharing (Salerno et al., 2017), storing (Stylianou and Talias, 2017; Larkin and Hystad, 2017; Kostkova et al., 2016; Conway, 2014; Gossec et al., 2020; O’Doherty et al., 2016), training AIS (Murphy et al., 2021) to destructing data (Wang et al., 2021).

The main harm of privacy breaches may be the risks of re-identification. Even if data are anonymized, many studies have shown that individuals can often be re-identified (Lee et al., 2016). Re-identification can be done by linking anonymous data, meta data (Gilbert et al., 2019) and datasets (Docherty and Lone, 2015) and is made easier with interoperable datasets (Delpierre and Kelly-Irving, 2018). Many authors in this review agree that re-identification risks are high with BD and related technologies. The re-identification risk increases with data’s dimensionality, i.e., the number of variables of data (e.g., age, location, weight, any other physiological trait, genetic information, etc.) (Bellazzi, 2014; Cahan et al., 2019; Mooney and Pejaver, 2018). Re-identification risks also increase with the low prevalence of the variable (e.g., rare medical conditions) (Docherty and Lone, 2015; Ballantyne, 2019; Demuro et al., 2020), the quantity of personal data in the public domain (Tsai and Junod, 2018), data linkage (Manrique de Lara and Pelaez-Ballestas, 2020), combination of data (Rennie et al., 2020), the improvement of data mining methods (Lee et al., 2016), and who has access to it, at the end, creating various degrees of de-identification (Ballantyne, 2019).

Surveillance activities raise particular concerns in terms of privacy. They are troubling considering the staggering amounts of data held by health organizations, corporations (Celedonia et al., 2021; Lodders and Paterson, 2020) and governments that can be used against the interest of individuals (Andanda, 2019; Baldassarre et al., 2020; Evans et al., 2020). The risk of surveillance is an unavoidable trade-off of the of BD (Ngan and Kelmenson, 2021; Howe and Elenberg, 2020) and AIS in health-related activities and one that attenuates its possible benefits (Galetsi et al., 2019; Mootz et al., 2020). For example, passive technologies such as imbedded sensors are less intrusive than direct observation (Grigorovich and Kontos, 2020), but nonetheless imply the collection of immense quantities of data. In the context of the COVID-19 pandemic, citizen populations were watched in order to prevent the spread of the disease, but many expressed concerns that this information be used for other purposes (Sarbadhikari and Pradhan, 2020; Shachar et al., 2020; Naudé, 2020; Shen and Wang, 2021).

The operationalization of privacy standards faces several challenges. It is not clear how to use the polysemic concept of privacy (Mooney and Pejaver, 2018; Conway, 2014; Casanovas et al., 2017; Snell, 2019). Some suggest to distinguish different forms of privacy, which certain forms are more at risk with BD such as informational privacy (Ienca et al., 2016) or physical privacy through surveillance (Bennett, 2019). The complexity may also arise because of the overlapping of privacy with a large spectrum of ethical values such as trust, transparency, security and property over who has access to the data and for what uses (Canaway et al., 2019). Contexts may also influence the definition and operationalization of privacy. For example, different areas of research have various methodologies and tools, complicating the protection of privacy in interdisciplinary health research (Casanovas et al., 2017).

Culture could also influence how privacy is understood, raising the question of whether a core definition should be used across all settings or not (Vayena et al., 2015). Also, in some political and economic contexts, citizens may consider that privacy concerns are irrelevant because of the level of surveillance already imposed by the State (Liu and Graham, 2021). Authors also note regularly the paradox between the perceived lack of concern of people toward sharing identifiable information on internet platforms (Aiello et al., 2020; Mikal et al., 2016; Yang and Chen, 2018; Snell, 2019; Young, 2018) and, at the same time, the fear of privacy breach related to participation in research project (de Graaf et al., 2015; Montgomery et al., 2018; Mootz et al., 2020; Tsai and Junod, 2018; Wongkoblap et al., 2017), public health interventions (O’Doherty et al., 2016) or any other health activities (Yang and Chen, 2018).

Common strategies have been proposed to protect privacy such as de-identification (Comess et al., 2020; Aebi et al., 2021), anonymization, (Comess et al., 2020; Gilbert et al., 2020; Aebi et al., 2021) and geo-masking (Comess et al., 2020; Aebi et al., 2021). However, these strategies face several limitations such as the complex language of privacy policies (Aiello et al., 2020), the lack of transparency about the protection mechanism used (Casanovas et al., 2017), the overall cost of the protection mechanisms (Zhang et al., 2017; Kayaalp, 2018), the use of protection mechanism more adequate for “small data” rather than BD (Wang and Alexander, 2020; Sun et al., 2020), and the ambiguous status of sensible data shared on social media (Celedonia et al., 2021; Gilbert et al., 2020).

The value of privacy conflicts with the possible benefits associated with using BD and AI in health-related contexts (Sparrow and Hatherley, 2019; Cool, 2016; Goodman, 2020; Salerno et al., 2017; Gilbert et al., 2019; van Heerden et al., 2020; Yeung, 2018; Igual et al., 2013). During the COVID pandemics, empirical data have shown that, for certain people, the loss of privacy was perceived as a trade-off for public health (Liu and Graham, 2021; Degeling et al., 2020). Aside from greater public health outcomes and prevention (Dolley, 2018; Alemayehu and Berger, 2016; Aiello et al., 2020; Horvitz and Mulligan, 2015; Roberts, 2019; Raza and Luheshi, 2016; Adkins, 2017; Mootz et al., 2020; Hodgson et al., 2020; Mahlmann et al., 2017), authors suggest that the promotion of scientific innovation could outweigh privacy (Heitmueller et al., 2014; Balas et al., 2015; Manrique de Lara and Pelaez-Ballestas, 2020; Salerno et al., 2017; Liu and Bressler, 2020; Comess et al., 2020; Terry, 2014; Wyllie and Davies, 2015).