Background: Clinicians’ scope of responsibilities is being steadily transformed by digital health solutions that operate with or without artificial intelligence (DAI solutions). Most tools developed to foster ethical practices lack rigor and do not concurrently capture the health, social, economic, and environmental issues that such solutions raise.
Objective: To support clinical leadership in this field, we aimed to develop a comprehensive, valid, and reliable tool that measures the responsibility of DAI solutions by adapting the multidimensional and already validated Responsible Innovation in Health Tool.
Methods: We conducted a 3-phase mixed methods study. Relying on a scoping review of available tools, phase 1 (concept mapping) led to a preliminary version of the Responsible DAI solutions Assessment Tool. In phase 2, an international 2-round e-Delphi expert panel rated on a 5-level scale the importance, clarity, and appropriateness of the tool’s components. In phase 3, a total of 2 raters independently applied the revised tool to a sample of DAI solutions (n=25), interrater reliability was measured, and final minor changes were made to the tool.
Results: The mapping process identified a comprehensive set of responsibility premises, screening criteria, and assessment attributes specific to DAI solutions. e-Delphi experts critically assessed these new components and provided comments to increase content validity (n=293), and after round 2, consensus was reached on 85% (22/26) of the items surveyed. Interrater agreement was substantial for a subcriterion and almost perfect for all other criteria and assessment attributes.
Conclusions: The Responsible DAI solutions Assessment Tool offers a comprehensive, valid, and reliable means of assessing the degree of responsibility of DAI solutions in health. As regulation remains limited, this forward-looking tool has the potential to change practice toward more equitable as well as economically and environmentally sustainable digital health care.
Over the past decade, digital health solutions and those relying on artificial intelligence (AI) have exponentially grown and expanded research and health care practices in ways that were previously unthinkable . As AI is entirely dependent on digital infrastructures [ ], the inclusive term “digital health solutions that operate with or without AI” (“DAI solutions”) is used throughout this paper to refer to electronic systems that rely on software and possibly also hardware to generate, store, or process data and that operate with or without AI [ ]. Although DAI solutions are steadily transforming health systems [ ] as well as clinicians’ practices and scope of responsibilities [ ], health care providers involved in the development and assessment of these tools mainly focus their attention on safety, effectiveness, and biases [ ]. However, as the DAI solutions industry within which the digital health field evolves is not guided by a professional care ethos, strong clinical leadership is required for DAI solutions in health to remain aligned not only with patients’ needs and health care values and goals [ ] but also with current knowledge on the effects of climate change on health [ , ]. Research shows that DAI solutions in health that are not properly designed or implemented increase digital health inequalities [ , ] and that their use requires more devices and data infrastructures that cause environmental harms (eg, material mining, e-waste disposal, and energy use) [ , ]. As “a relentless drive” to use larger amounts of data and more sophisticated computational capacities comes with higher environmental costs, the powerful clinical tools that DAI solutions offer thus entail substantial “trade-offs” that clinicians can no longer ignore [ ].
As health care providers and health systems worldwide will be on the front line tackling the health effects of climate change and growing social and economic disparities [, ], clinical leaders (eg, physicians, nurses, occupational therapists, and psychologists) will have to play a much broader role in the design and assessment of DAI solutions. They must be able to anticipate and properly prepare trainees and practitioners to address the health, social, economic, and environmental impacts of the DAI solutions they work with or recommend to patients or that their organizations acquire. As these multidimensional impacts are linked to each other, clinical leaders urgently require tools so they can comprehensively and efficiently assess the relevance of DAI solutions “prior to implementation” and “lead the change” needed in partnership with other health care stakeholders for such solutions to support patient care and health systems in a meaningful and responsible way [ ].
Many ethical principles (eg, privacy, accountability, and robustness) have been proposed to foster responsibility in the digital industry, either specifically for health care  or for multiple sectors [ ] and either for digital solutions [ ] or for AI [ ]. The scoping review our team recently completed highlights key gaps in the practice-oriented tools developed since 2015 [ ]. First, these tools are highly heterogeneous, which may facilitate “mixing and matching” [ ] principles that do not rely on a solidly defined conceptual framework. For instance, among the 56 tools we identified, ≥50% (≥10/19) of those from the health sector relied on a small number of principles (n=10), ≥50% (≥20/37) of the multisector tools covered twice as many principles (n=19), and most tools (≥29/56, ≥50%) disregarded 21 principles over a total of 40 principles found in the 56 tools. Second, the methodology used to develop 82% of the tools was not defined, 18% used engagement methods (eg, workshops and consultations), and none reported how quality was assessed. This is a major research gap as tools that lack a solid methodology may undermine at its roots the very goal of fostering responsible DAI solutions: clinical leaders are unlikely to adopt them “if their quality or credibility is perceived as low” [ ].
Goal of the Study and Approach
To support clinical leadership in this rapidly evolving field, the goal of our study was to develop a comprehensive, valid, and reliable tool to measure the degree of responsibility of DAI solutions in health. The Responsible Innovation in Health (RIH) framework , which brings together key health, social, economic, and environmental issues, offered a solid basis to develop such a tool as RIH is anchored in an evidence-informed health research tradition [ ]. Its accompanying RIH assessment tool is one of the rare tools in the field of responsible research and innovation that is specific to the health care sector [ ] and that provides a conceptually valid [ ] and reliable [ ] quantitative measure of responsibility. However, the RIH tool does not capture responsibility issues specific to DAI solutions (eg, data management).
Following Stilgoe et al , for whom responsible innovation means “taking care of the future through collective stewardship of science and innovation in the present,” RIH is forward-looking, pragmatic, and multidisciplinary [ ]. RIH goes beyond deontology and bioethics as it aims to steer health innovation toward equitable as well as economically and environmentally sustainable health systems [ ]. The RIH framework approaches responsibility as a matter of degree, which can be appraised by examining 9 responsibility attributes falling within five value domains: (1) population health value (health relevance; health inequalities; and ethical, legal, and social issues), (2) health system value (responsiveness, inclusiveness, and level and intensity of care), (3) economic value (frugality), (4) organizational value (business model), and (5) environmental value (eco-responsibility). Although the first 2 value domains are familiar to clinicians, the other 3 offer key considerations for taking care of the future when developing health innovations [ ]. Through its “Frugality” attribute, RIH underscores that an innovation adds economic value when it is designed to be affordable and easy to use and optimized for its context of use without neglecting low-resource settings [ ]. The “Business model” attribute emphasizes organizations that are stakeholder centered (ie, that create value for society, not only for shareholders [ , ]). Finally, the “Eco-responsibility” attribute recognizes that planetary health and human health are deeply intertwined [ ].
In this study, the iterative research process that led to the RIH tool [, ] was replicated to (1) adapt the RIH tool to the specificities of DAI solutions, (2) validate the constructs of the resulting Responsible DAI solutions Assessment Tool (hereafter referred to as the “tool”), and (3) assess its reliability. To facilitate readers’ understanding, summarizes the tool’s key components: 4 premises, 5 screening criteria, and 14 assessment attributes. The type of information its 3-step application process (screening, assessment, and scoring) requires is described in the tool, which can be found in . The attributes use a 4-level scale ranging from A to D, where A implies a “high degree of responsibility” and D implies “no particular signs of responsibility.” As the attributes do not measure “irresponsibility,” the screening criteria constitute baseline responsibility requirements (eg, efficacy, safety, and privacy) that serve as a “stopping rule” in the assessment process.
To achieve our study’s goal, we conducted a 3-phase mixed methods study, which is now mainstream in health services research . It offers “an overarching methodological framework to a multiyear project” where the ability to build on what was learned previously is paramount [ ]. Its purpose “is to address a set of incremental research questions” that all advance a broader research goal [ ]. As shows, the 3 phases—concept mapping, content validity assessment, and interrater reliability assessment—were sequentially aligned to iteratively collect, analyze, and combine the quantitative and qualitative data needed for each incremental research objective (described in the following sections). Throughout the study, we placed a greater emphasis on quantitative methods as this is recommended when qualitative data supplement the development of robust instruments [ ]. Our study is reported following the Mixed Methods Research checklist [ ] (the study protocol is available elsewhere [ ]).
Phase 1: Concept Mapping
The objective of phase 1 was to identify responsibility principles and best practices specific to DAI solutions missing from the original RIH tool. This phase relied on our scoping review  and concept mapping, which refers to a “structured process” that gathers “input from multiple participants” and uses qualitative pattern matching and quantitative multivariate analyses to produce an exhaustive map of a conceptual domain [ ]. The 3-step process we followed to generate, structure, and represent “as completely as possible all of the key facets” [ ] of responsibility in DAI health solutions (ie, the conceptual domain of interest in our study) is fully described in [ , - , , - , - ].
Following an accountable qualitative thematic analysis strategy , LR and RRO categorized each principle found in the 56 tools included in the scoping review using the definitions provided by their authors. The quantitative analyses first examined the distribution of the principles found across the tools, which shed light on the responsibility constructs that they prioritized or disregarded. We then stratified the analyses along 3 subsets of tools—those from academia, governments, and the business sector—and applied a network analysis [ ]. The aim was to examine the connection patterns between the tools and the principles they promoted (ie, “Principle A” is linked to “Tool 1” when the latter relies on that principle). By calculating the degree of centrality (in-degree), which represents the proportion of connections that a principle has compared with all possible connections it may have with the other tools in the subset, we obtained a ranking order in which more than one principle could occupy the same position. This facilitated a systematic comparison of the responsibility constructs that these tools sought to operationalize through questions, recommendations, criteria, and “dos and don’ts,” among other things.
Then, 4 researchers (LR, RRO, PL, and an AI ethics expert) independently mapped all principles across the RIH tool’s components: (1) premises (how responsibility is defined for the purposes of the tool), (2) screening criteria (baseline responsibility requirements), and (3) assessment attributes (degree to which responsibility characteristics are present). Each team member examined what type of revision was needed to capture the responsibility principles specific to DAI solutions (eg, modifying an existing premise, eliminating it, or adding a new one). Finally, we identified through team deliberations adaptations that covered the principles already captured in the RIH tool (eg, revising the “Ethical, legal, and social issues” attribute to cover specific data-related consent and compliance issues), those that could be aggregated (eg, antidiscrimination and fairness), and those not captured in the RIH tool that called for new attributes (eg, interoperability and data governance).
Phase 2: Content Validity Assessment
The objective of phase 2 was to ensure the tool’s content validity, which refers to “the degree to which the content of an instrument is an adequate reflection of the construct to be measured” . It relied on an international 2-round e-Delphi exercise [ ]. Delphi research techniques are often used in emerging clinical areas of practice that “span multiple areas of expertise” and where consensual guidelines are lacking [ ]. For a multidisciplinary panel of international experts to critically assess and improve the content validity [ ] of the first version of the tool, the round 1 survey of our e-Delphi exercise comprised 22 closed-ended items using a 5-level Likert scale and 22 free-text boxes for experts to explain their ratings (excluding the research ethics consent form and demographic survey items). As the original RIH components had already been validated, the round 1 survey focused on the importance, clarity, applicability, and comprehensiveness of the new components specific to DAI solutions (indicated in ). On the basis of the results of round 1, a revised version of the tool was developed for round 2, which comprised 20 closed-ended items and 20 free-text boxes. It excluded items for which consensus had been reached and introduced the scales associated with each attribute (the surveys can be found in ). After each round, personalized feedback with individual responses and measures of central tendency, as well as the panel comments, was sent to each expert [ ].
Three measures had to be met to reach consensus: (1) at least 51% of experts scoring the item on the 2 highest levels (4 and 5), (2) an SD of ≤1.5, and (3) an IQR of ≤1.0 . We applied a rigorous qualitative thematic approach [ ] to analyze free-text responses. In total, 3 researchers (LR, RRO, and PL) independently categorized the comments, made proposals to address them, and then agreed on the changes required to improve the tool.
Phase 3: Interrater Reliability Assessment
The objective of phase 3 was to assess the reliability of the tool by measuring interrater agreement and suggest measurement revisions if needed . Interrater reliability refers to the extent to which 2 or more raters classify the same set of objects in the same way [ ]. Following the recommendations by Gwet [ ] on the number of objects required to achieve a sufficient level of accuracy and minimize the percentage of agreement SE, an error margin of –0.20 to +0.20 was used to determine our sample size, that is, 25 DAI solutions. We first identified 45 real-world solutions, gathered publicly available information about them, and proceeded in a stepwise fashion to create a balanced and diversified sample. We selected solutions operating with or without AI, pursuing different purposes (eg, self-management, diagnosis, treatment, and administration), developed by diverse organizations (for-profit, not-for-profit, governmental, and nongovernmental organizations), and used in different contexts of care and regions. For the 2 raters (RRO and LR) to apply the tool as intended, we searched each solution’s website to collect information addressing the tool’s criteria and attributes (terms of reference, privacy or sustainability policy, user guides, governance structure, and annual reports). We tabulated relevant excerpts from all 25 DAI solutions in a Microsoft Excel (Microsoft Corp) “scorecard” that both raters completed independently. As start-ups tended to share less detailed documentation than large firms, PL adapted the content found on other developers’ websites for the scorecard to contain all the information needed to score each criterion and attribute for all 25 solutions ( ).
Once each rater had independently completed their assessment, we calculated (1) a nonadjusted index (percentage of agreement), (2) a more paradox-resistant chance-adjusted index (the Gwet agreement coefficient), (3) SEs, (4) 95% CIs, and (5) P values . The interpretation of the strength of the Gwet agreement coefficient, where 1 represents maximum reliability and 0 represents no reliability, follows the Landis-Koch scale [ ]: poor (<0.0), slight (0.0-0.20), fair (0.21-0.40), moderate (0.41-0.60), substantial (0.61-0.80), and almost perfect (0.81-1.00). Finally, a fourth team member (HPS) chaired a meeting for the 2 raters to deliberate over diverging scores, reach consensus, and identify final minor improvements to the tool.
Ethics approval was obtained from the Health Sciences Research Ethics Review Board of the Université de Montréal (CERSES-20-144-D).
Phase 1: Tool Comprehensiveness
Although the scoping review data set is available elsewhere , the databases used as well as the inclusion and exclusion criteria are described in . [ ] summarizes the selection process following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) [ ]. We retained a total of 56 tools, 12 (21%) from academic literature and 44 (79%) from gray literature.
illustrates the results of the mapping process that led to the first version of the tool. It shows how the 40 principles identified after systematically coding each tool are linked to the tool’s premises, inclusion and exclusion criteria, and assessment attributes. After team deliberation, we revised 2 RIH premises and introduced 2 new ones (“AI for good is not automatically responsible” and “Relevance of digitalization”). We revised all RIH inclusion criteria and added a new exclusion criterion that covered 4 areas considered particularly problematic in DAI solutions (“Data reselling as the primary business model,” “Deliberately deceptive solution,” “Lack of cybersecurity and personal data protection,” and “AI relying on biased datasets”). We revised existing RIH attributes and integrated 3 new attributes: “Human agency,” “Interoperability,” and “Data governance.” Finally, we adapted the RIH frugality and eco-responsibility attributes to account for both software and hardware that may be required to operate a DAI solution.
Phase 2: Tool Content Validity
shows the flowchart of the e-Delphi expert solicitation process, which began in April 2022. The information sources used to identify experts in 4 main disciplinary fields—health sciences, engineering and computer sciences, social sciences, and administration and law—are described in . A total of 799 invitations were sent to authors of documents screened during the phase 1 scoping review (n=755, 94.5%) and to purposefully sampled experts (n=44, 5.5%). After 3 reminders and excluding surveys not fully completed, 26 experts participated in the round 1 survey, with 224 comments and a 3.3% (26/799) response rate. Between June 2022 and October 2022, a total of 14 experts completed the round 2 survey, with 49 comments and a 50% (14/28) response rate.
describes the final panel composition, which included a similar proportion of men and women (13/26, 50% and 12/26, 46% in round 1 and 6/14, 43% and 6/14, 43% in round 2, respectively). A well-balanced representation across the 4 disciplinary fields was observed in round 1 (ranging from 5/26, 19% to 7/26, 27%). In round 2, a higher participation of social scientists (6/14, 43%) and health scientists (4/14, 29%) was observed, whereas a similar proportion (2/14, 14%) of engineers and computer scientists and of administration and law experts completed the survey. Most experts had >10 years of experience (17/26, 65% in round 1 and 9/14, 64% in round 2) and were employed in higher education institutions (21/26, 81% in round 1 and 12/14, 86% in round 2) in North America (16/26, 62% in round 1 and 7/14, 50% in round 2).
presents the results of the e-Delphi round 1 and round 2 surveys. In round 1, consensus was reached on 27% (6/22) of the survey items pertaining to the first version of the tool: the importance of 1 premise (“Context of use”), the applicability of 1 screening criterion (“DAI solution definition”), the importance of the “Human agency” attribute, the importance and clarity of the “Data governance” attribute, and the clarity of the “Programming and software eco-responsibility” attribute. On the basis of the comments received, which can be found in , we withdrew 1 premise (“AI for good”) and 1 screening criterion (“Relevance of digitalization”); formulated 1 new premise (“DAI solutions should tangibly improve current processes and means”); revised all remaining premises, criteria, and attributes; and developed the scales for all attributes. In round 2, consensus was reached on 80% (16/20) of the items surveyed for the second version of the tool. Experts agreed on the importance and clarity of all premises except for the clarity of “DAI solutions affect the determinants of health.” They agreed on the applicability of 1 screening criterion (“GA stage not reached”) but not on the applicability of “Nondisclosure of key DAI risks.” Consensus was reached on the clarity, importance, and appropriateness of the scales of all assessment attributes except for the clarity of “Human-centered interoperability” and for the appropriateness of the scale of “Programming and software eco-responsibility.” Overall, the content validity of 85% (22/26) of the items surveyed was confirmed after round 2. The comments received enabled our team to generate a third version of the tool that addressed all round 2 experts’ criticisms (see our responses in ).
|Characteristic||Round 1 (n=26), n (%)||Round 2 (n=14), n (%)|
|Woman (including cisgender and transgender)||12 (46)||6 (43)|
|Man (including cisgender and transgender)||13 (50)||6 (43)|
|Undisclosed or no response||1 (4)||2 (14)|
|Health sciences or public health||5 (19)||4 (29)|
|Engineering, computer sciences or data sciences, design, natural sciences, mathematics, statistics, or operational research||5 (19)||2 (14)|
|Social sciences or humanities||7 (27)||6 (43)|
|Business, public administration or management, and law or accounting||6 (23)||2 (14)|
|Multidisciplinary field||3 (12)||0 (0)|
|Years of experience|
|<5||4 (15)||1 (7)|
|5-10||5 (19)||4 (29)|
|>10||17 (65)||9 (64)|
|Higher education||21 (81)||12 (86)|
|For-profit organization, consultant firm, or privately funded research institution||3 (12)||1 (7)|
|Government or arm’s length public administration agency||1 (4)||0 (0)|
|Health care facility||1 (4)||1 (7)|
|North America||16 (62)||7 (50)|
|Europe||7 (27)||4 (29)|
|Asia||1 (4)||1 (7)|
|Oceania||2 (8)||2 (14)|
|Survey item||Score of 4 to 5, n (%)||Values, SD||Values, IQR||Consensus status|
|Round 1||Round 2||Round 1||Round 2||Round 1||Round 2||Round 1||Round 2|
|Premises (shortened for clarity)|
|Context of use|
|Clarity||12 (46)||10 (100)||1.2||0.4||1.8||0||✓|
|Importance||18 (69)||9 (90)||1.1||0.7||2.0||0||✓|
|Clarity||17 (65)||8 (80)||1.1||1.2||2.0||1.0||✓|
|AIc for good (withdrawn in round 2)|
|Digital determinants of health|
|Importance||17 (65)||8 (89)||1.5||0.7||2.0||0||✓|
|Clarity||19 (73)||7 (70)||1.0||1.4||1.8||1.8|
|Screening step inclusion and exclusion criteria|
|Relevance of digitalization (withdrawn in round 2)|
|GAe stage not reached|
|Applicability||13 (50)||13 (93)||1.0||0.6||2.0||1.0||✓|
|Nondisclosure of key DAI solutions risks|
|Applicability||17 (65)||10 (71)||1.2||1.2||2.0||1.8|
|Assessment step attributes|
|Clarity||18 (69)||10 (83)||1.0||0.9||2.0||0.3||✓|
|Appropriate scale||N/A||11 (85)||N/A||0.8||N/A||1.0||N/A||✓|
|Human-centered interoperability (round 1); care-centric interoperability (round 2)|
|Importance||19 (73)||11 (100)||1.0||0.5||1.8||0.5||✓|
|Clarity||18 (69)||9 (75)||1.0||1.0||2.0||1.3|
|Appropriate scale||N/A||9 (82)||N/A||0.8||N/A||1.0||N/A||✓|
|Importance||19 (73)||10 (91)||0.8||0.9||1.8||1.0||✓|
|Clarity||18 (69)||11 (92)||0.9||0.9||2.0||1.0||✓|
|Appropriate scale||N/A||9 (82)||N/A||1.0||N/A||1.0||N/A||✓|
|Appropriate scale||N/A||11 (92)||N/A||0.9||N/A||1.0||N/A||✓|
|Programming and software eco-responsibility|
|Importance||18 (69)||10 (83)||1.0||1.2||2.0||1.0||✓|
|Appropriate scale||N/A||9 (75)||N/A||1.0||N/A||1.3||N/A|
aThe survey items were formulated as follows: How important is this premise/criterion/attribute? How applicable is this criterion? Is this premise/attribute clearly defined? Is the scale appropriate?
bN/A: not applicable.
cAI: artificial intelligence.
dDAI solutions: digital health solutions that operate with or without artificial intelligence.
eGA: general availability.
Phase 3: Tool Reliability
A description of the 25 DAI health solutions selected for assessing the tool’s reliability can be found in. The sample comprised 52% (13/25) of solutions that operated with AI (eg, a wayfinding app for persons living with cognitive or physical impairment combining GPS technology and AI and an AI-based diabetic retinopathy screening system) and 48% (12/25) that operated without AI (eg, a platform to develop customized apps for health care facilities with limited digital infrastructures and a virtual reality–based treatment for individuals living with chronic lower back pain). An equal number of solutions (5/25, 20%) supported prevention, self-care, diagnostics, treatment, or administration. In total, 32% (8/25) of the solutions were designed to be used in a clinical setting only, 52% (13/25) were designed to be used in a nonclinical setting only, and 16% (4/25) were designed to be used in both settings. A total of 12% (3/25) of the solutions were developed by governmental agencies or user-led associations, 32% (8/25) were developed by not-for-profit organizations (universities and nongovernmental organizations), and 56% (14/25) were developed by for-profit organizations. According to their developers’ websites, 56% (14/25) of the solutions were in use in more than one continent.
shows the results of the interrater reliability assessment (the data set is available in ). For screening criteria, an almost perfect agreement was found for “DAI solution,” for 2 subcriteria of “Nondisclosure of DAI risks” applicable to all DAI solutions, and for “GA stage not reached.” A “substantial agreement” was obtained for the “Nondisclosure of DAI risks” subcriterion applicable only to AI solutions. As we first reached a “moderate agreement” for “Human agency” (results can be found in ), we decided to revise its definition and perform a second interrater agreement. This was aligned with our objective of suggesting measurement revisions if needed. The reliability of the definitive version of the tool was high as an “almost perfect” agreement was obtained for all assessment attributes.
|Survey item||Agreement (%), SE (95% CI)||DAI solutionsb (%)||P value||Gwet AC1c or AC2d coefficient (SE; 95% CI)||P value||Interpretation|
|DAI solution definition||0 (1-1)||100||N/Ae||1 (0; 1-1)||N/A||Almost perfect|
|Nondisclosure of key risks||0 (1-1)||100||N/A||1 (0; 1-1)||N/A||Almost perfect|
|Nondisclosure of key risks (applicable to AIf only)||0.10415 (0.619-1)||85||<.001||0.792 (0.1602; 0.443-1)||<.001||Substantial|
|GAg stage not reached||0 (1-1)||100||N/A||1 (0; 1-1)||N/A||Almost perfect|
|Human agency (revised)||0.01021 (0.966-1)||99||<.001||0.964 (0.02946; 0.903-1)||<.001||Almost perfect|
|Care-centric interoperability||0.02408 (0.905-1)||96||<.001||0.865 (0.07190; 0.717-1)||<.001||Almost perfect|
|Software frugality||0.01992 (0.899-0.981)||94||<.001||0.837 (0.05622; 0.721-0.953)||<.001||Almost perfect|
|Data governance||0.02552 (0.897-1)||95||<.001||0.881 (0.06450: 0.748-1)||<.001||Almost perfect|
|Programming and software eco-responsibility||0.00245 (0.992-1)||100||<.001||0.994 (0.00594; 0.982-1)||<.001||Almost perfect|
aGwet first-order agreement coefficient is shown for the nominal ratings of the screening criteria (yes or no), and Gwet second-order agreement coefficient is shown for the ordinal ratings of the assessment attributes (A, B, C, and D). We used unweighted coefficients for nominal ratings and weighted coefficients for ordinal ratings (using quadratic weights). Interpretation follows the Landis-Koch scale: 0.8 to 1=almost perfect; 0.6 to 0.8=substantial; 0.4 to 0.6=moderate; 0.2 to 0.4=fair; 0 to 0.2=slight; and <0=poor . The results of the first interrater agreement for “Human agency” as well as the changes made to this attribute can be found in .
bDAI solutions: digital health solutions that operate with or without artificial intelligence.
cAC1: first-order agreement coefficient.
dAC2: second-order agreement coefficient.
eN/A: not applicable.
fAI: artificial intelligence.
gGA: general availability.
Principal Findings and Comparison With Prior Work
Considering that the current biggest challenges to health are at the interface of climate change and growing inequalities  and that the carbon footprint of digital services is increasing by 8% annually [ ], this study’s contributions are 3-fold.
Informing the Responsible Design and Adoption of DAI Solutions
First, the Responsible DAI solutions Assessment Tool is among the first tools to offer a comprehensive, valid, and reliable means to measure the degree of responsibility of DAI health solutions that can be applied by clinicians and other health innovation stakeholders. On the one hand, it can inform “supply side” decisions made by those who design DAI solutions, such as data scientists, programmers, clinical investigators, entrepreneurs, investors, research funders, and incubators. On the other hand, it can inform “demand side” decisions, including those of purchasers, implementers, patients, clinicians, and health care managers (see the “Who can apply the Tool and how?” section in). Although the RIH framework and tool have been used to analyze responsibility challenges of DAI solutions [ , , ], scholars, clinicians, and decision makers have been calling for a concise tool that could also account for issues specific to DAI health solutions [ , ].
Screening of Baseline Responsibility Requirements
Second, the tool’s 3-step application process enables clinicians to swiftly screen whether a DAI solution lacks baseline responsibility requirements before proceeding to a full assessment. For instance, 1 of the 5 screening criteria requires documenting whether “the DAI solution has been proven effective and safe to human health” by using publicly available evidence such as peer-reviewed scientific articles or reports by regulatory agencies (see the “Sources of information to look for before applying the Tool” section in). Here, the assumption is that, if “must-have” requirements such as safety and effectiveness are not met, the solution cannot be considered responsible and, thus, there is little value in further assessing the extent to which responsibility attributes may or may not be fulfilled (unless the intent is to use the tool to improve the solution; see the following sections). A similar logic applies to the exclusion criterion “Nondisclosure of key DAI risks,” but in this case, information sources that may be used to apply the tool are those made publicly available by solution developers (eg, terms and conditions statement, data protection, and privacy policies). This criterion examines whether the organization that makes the DAI solution available to end users refrains from selling user-related data [ ]; makes explicit its cybersecurity, privacy, and personal data protection measures; and clearly communicates how potential biases in the data set used to train an AI were mitigated (when applicable) [ ]. Acknowledging that such information sources are of lower quality, the tool nonetheless strongly encourages solution developers (ie, data scientists, programmers, entrepreneurs, and high-level executives) to make their commitments to responsible DAI solutions explicit and, thus, accountable [ ]. This seems particularly important as “patients and clinicians struggle to select digital health tools in an environment with inconsistent regulation and sparse information” on their risks; benefits; and ethical, legal, and social issues [ ].
An Integrated Set of RIH Attributes Specific to DAI Solutions: From Human Agency to Eco-Responsibility
Third, the tool’s new attributes and their descriptive mutually exclusive scales can help clinicians identify and compare the degree of responsibility of different DAI solutions. For Obermeyer and Topol , the technical choices and human values underpinning the training of AI can either “scale up” biases based on socially determined characteristics such as race and gender or help “fight against” them. The “Human agency” attribute provides further practical guidance as its scale describes 4 concrete agency enablers that a DAI solution can proactively embed in its design and use. These enablers should help clinicians and patients (1) understand the measures, recommendations, decisions, or outputs of a DAI solution (eg, data visualization and transparency if an AI-based solution is unexplainable [ ]); (2) discuss their implications with managers or staff when needed (eg, dedicated point of service); (3) act in accordance with their own goals without undue pressure (eg, freedom to override an AI-based decision [ ]); and (4) have their concerns acted upon through an appeal, audit, review, or redress mechanism (eg, ombudsman [ ]). This new attribute is aligned with recent efforts to define the “minimum information” required for users to better understand the “intended predictions, target populations, and hidden biases” of DAI solutions (see the Minimum Information for Medical AI Reporting) [ ]. It also supplements other key RIH attributes in striving to reduce avoidable health status differences across individuals and groups (“Health inequalities”), avoid user parameters that preclude legal rights to be exercised (“Mitigation of ELSIs”), and overcome a poor understanding of different users’ varying needs (“Inclusiveness”) [ , ].
The “Care-centric interoperability” attribute refers to how smoothly a DAI solution can securely operate within and across clinical and nonclinical settings without adding cognitive or administrative burden to users . It is based on a broader understanding of the interoperability standards promoted for a safe integration and interfacing of digital and nondigital devices in a health system [ ]. Its scale stresses four characteristics that can be embedded in a DAI solution design: (1) aligning the solution with its users’ data management practices (and not vice versa) to minimize cognitive and administrative burden, (2) aligning the solution with its users’ digital infrastructures (eg, operable on widely available systems and devices [ , ]), (3) incorporating data sharing functionalities that “follow the patient” along clinical pathways or practitioners’ work processes (eg, nonproprietary software and data portability [ ]), and (4) ensuring that it can securely evolve with users’ digital infrastructures (eg, built-in security features in software as a service and auditable logs [ ]).
The “Software frugality” attribute refers to the ability to deliver greater value to more people by using fewer resources, such as capital, materials, energy, and labor time . Frugal innovation may be easily overlooked in the health care sector, but it clearly matters to the future of health systems [ , ]. Grounded in an up-to-date scientific understanding, this attribute recognizes that frugality is not about creating “the cheapest products” [ ]; rather, it is about increasing their economic value by designing high-quality solutions that are affordable and usable and fit with their context of use. The scale of this attribute stresses that responsible software should meet three frugal innovation characteristics [ ]: (1) affordability (which may result from software development strategies; open-source programming tools; or low technical support, update, and maintenance needs [ ]), (2) focus on core user-facing functionalities that meet a larger number of user capabilities (eg, universal interface design for users with low literacy), and (3) maximized fit between functionalities and user location–dependent digital capacities [ ] (eg, edge computing for settings where connectivity is compromised).
The “Data governance” attribute responds to a widely shared consensus among scientific  and policy [ ] communities for proper oversight of data. It refers to the stewardship, structures, and processes that an organization sets in place to ensure full control over the entire data life cycle. The scale brings forward four mechanisms that can be combined to support responsible data governance: (1) a chief data officer or committee accountable for the way employees gather, exploit, generate, store, share (voluntarily or not), or destroy data and for any data-related breaches or incidents [ ]; (2) a training program for managers and employees to remain up-to-date and properly skilled in data management; (3) data protection practices relying on performance indicators or standards (eg, ISO/IEC 27001: information security management and ISO/TS 82304-2: quality and reliability of health and wellness apps [ ]); and (4) an auditable data governance reporting system [ , ].
Finally, 2 distinct attributes were created to fully capture the environmental harms arising from hardware on the one hand and from programming and software on the other. “Programming and software eco-responsibility” refers to a product, process, or method that reduces as much as possible the negative environmental impacts. It spans the use of clean energy sources and the reduction of the energy consumed when developing AI and software and archiving data. The scale of this attribute highlights three eco-responsible practices: (1) choosing programming, modeling, or computational techniques that substantially reduce the quantity of energy and time required to develop a DAI solution (eg, TinyML); (2) using highly energy-efficient central processing units and computers; and (3) selecting data centers and server farms where greenhouse gas emissions are reduced to a minimum (net zero) or where more greenhouse gases are removed from the atmosphere than emitted (climate positive) [, ]. This adaptation of the RIH tool responds to current knowledge on the environmental impacts of the ever-increasing energy demands of complex data computational practices, including the training of algorithms [ ], and of the growing use of digital devices that consume rare-earth metals and have harmful end-of-life disposal outcomes [ ]. Given the numerous hardware components that may surround a DAI solution, the tool clearly indicates whether the attributes “Hardware frugality” and “Hardware eco-responsibility” apply (see the “Scope of the assessment” section in ). Hardware equipment should be included in the assessment when its raison d’être is to support the DAI solution and it is part of the minimal requirements for the solution to deliver its service. For instance, a finger sensor used to record an ECG using a smartphone fulfills these 2 criteria but not the smartphone. Similarly, surgical robot hardware components meet the 2 criteria (their raison d’être is to support the surgical procedure, and the latter would not be possible without them) [ ]. Attending to hardware eco-responsibility concerns implies reducing environmental harms at key stages in a product’s life cycle, which include (1) raw material sourcing (eg, free of substances that are harmful and toxic to ecosystems), (2) manufacturing (eg, compliance with national or international environmental regulations), (3) distribution (eg, packaging and transportation), (4) use (eg, durability and repairability), and (5) disposal (eg, designed to be recycled, disassembled, remanufactured, composted, or biologically degraded) [ ].
Implications for Practice
This concise yet comprehensive forward-looking tool is not without limitations, but it has the potential to change both thinking and practice in the rapidly evolving field of DAI health solutions. These solutions may drive many improvements in health care . Nevertheless, the pace at which they are being developed remains unprecedented when compared with other medical advances such as minimally invasive surgery, interventional radiology, or genomics [ ]. Although many scholars underscore that DAI solutions should be used “in compliance with relevant laws” [ ], regulatory frameworks remain scant, and policy progresses are slow [ ]. Current regulatory and policy limitations and the lack of robust assessment tools put patients, clinicians, and health care managers at risk [ ] not only of biases but also of diversion from health systems’ key mission: improving health in an economically and environmentally sustainable way [ ]. As many decisions driving the supply of DAI solutions are made outside the health sector [ ], clinicians are currently largely unequipped to anticipate and handle their health, social, economic, and environmental impacts [ ].
The tool was specifically designed to support clinicians in the broader role they should play as “change agents”  in the digital health field. Thanks to a multidisciplinary expert panel, the tool’s attributes are clearly defined, and its scales describe key responsibility enablers, characteristics, or mechanisms in a tangible way. Its practical value lies in the fact that it can be applied in two distinct ways: (1) as a formal evidence-informed assessment tool to measure the degree of responsibility of a DAI solution or (2) as a design or procurement brief (or template) to explore the suitability of a given DAI solution for patient care and clinical practice and guide its development, acquisition, implementation, or use. In both situations, the overall responsibility score is considered invalid if the screening criteria are not met. When used as a formal assessment tool, specific steps should be followed for the tool to deliver a valid score—after having searched, retrieved, and critically analyzed sources of information pertaining to each screening criterion and assessment attribute, an interdisciplinary team (2-5 raters with research skills) must first apply the tool independently and then reach consensus. As described in , when disagreements between raters are found, the team should deliberate to identify potential errors or misunderstandings. The consensus score should neither be “forced” nor “averaged”—it should establish a strong correspondence between the information available and the question (for the screening criteria) or the scale item (for the assessment attributes) formulated in the tool.
Of course, one of the tool’s limitations lies in the information sources required to rate each criterion and attribute. Although the strongest sources of evidence remain independent peer-reviewed publications, few are likely to be available for an emerging DAI solution . Moreover, the solution and the organization that makes it available to users may change over a brief period (eg, acquisitions of start-ups are frequent in the digital technology industry) [ ]. The scope of these changes may significantly affect the adequacy of the scientific evidence available and the degree of responsibility of the solution. Therefore, those who apply the tool should remain critical of the information provided by developers and reconduct the assessment whenever significant changes are made to the solution or organization. As novel applications of AI keep emerging, such as generative AI that uses natural language processing to create textual content (eg, ChatGPT), it will be important to keep abreast of technological advances and apply the tool rigorously (ie, as described in ). Its definition of responsibility is anchored in the RIH scholarship, which largely differs from definitions found under the “responsible AI” umbrella term [ ].
There are 3 limitations to this study that are partially mitigated by the strengths of a mixed methods study design . When launching phase 1, we were challenged by the velocity at which tools to foster responsibility in DAI solutions had been developed (ranging from 3 in 2016 to 25 in 2020), and their quantity precluded an in-depth qualitative analysis of their conceptual overlaps. However, as 93% of the principles came from tools published before 2019, we are confident that significant principles have not been omitted [ ]. In phase 2, we could not gather information about experts who ignored the invitation (690/799, 86.4%), did not click on “participate” (63/799, 7.9%), or did not complete the round 1 survey (14/799, 1.8%). However, the final panel size is adequate for an e-Delphi study, a high participation rate in round 2 (14/28, 50%) increases internal validity, and using 3 concurrent measures to determine consensus exceeds standards often seen in such studies [ ]. In phase 3, objects and raters were not randomly selected, which limits the ability to draw inferences. The tool’s reliability is predicated on having raters sufficiently familiar with its premises, criteria, and attributes. An overarching strength of this 3-phase study was to have built on the scientific groundwork that led to the RIH tool [ , ]. Thus, our team had a good command of the methods needed for “measuring the constructs of primary interest” [ ].
Clinicians active in research have made great strides to work with DAI solution developers to address key clinical issues [, , ], and ground-breaking scholarly and policy work has brought to light the numerous ethical concerns that arise with their development and use [ , , , ]. However, tools developed to foster responsibility in DAI solutions focus on fragmented sets of principles, rarely offer measurable indicators, and lack methodological rigor [ ]. Thus, we applied a rigorous study design to deliver a rigorous tool. Further actions include actively disseminating the tool through our research collaborators and developing multimedia materials to support its use [ ]. Although strong clinical leadership is required for high-quality digital health care to materialize in practice, the Responsible DAI solutions Assessment Tool can help clinical leaders contribute to the design and use of DAI solutions with a high degree of responsibility. It offers a comprehensive, valid, and reliable means to help steer DAI solutions toward equitable as well as economically and environmentally sustainable digital health care.
This study was funded through a peer-reviewed call for proposals by the International Observatory on the Societal Impacts of Artificial Intelligence and Digital Technology. Our research team benefits from an operating grant from the Canadian Institutes of Health Research (FDN-143294). Our research center is supported by the Fonds de la recherche du Québec–Santé. The funders of our study did not play any role in its design, conduct, or reporting. The authors would like to thank the research collaborators who provided them with insightful comments throughout the study: Cécile Petitgand, Barbara Decelle, Jean-Louis Denis, Alison Marchildon, Margie Mendel, Catherine Régis, and Philippe Després. Gabrielle Veilleux-Verreault performed key knowledge synthesis tasks, shared important artificial intelligence ethics insights, and programmed the e-Delphi exercise platform (Mésydel). This study would not have been possible without the precious insights and dedication of the international experts who participated in our phase 2 e-Delphi exercise.
The data sets collected and analyzed are available in the body of the text or in.
All authors confirm that they had full access to all the data in the study and accept responsibility for submission for publication. They all meet the four criteria for authorship in the International Committee of Medical Journal Editors recommendations: (1) substantial contributions to the conception or design of the work or the acquisition, analysis, or interpretation of data for the work; (2) drafting the work or revising it critically for important intellectual content; (3) final approval of the version to be published; and (4) agreement to be accountable for all aspects of the work ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. PL contributed to conceptualization, methodology, supervision, formal analysis, writing (original draft, review, and editing), project administration, and funding acquisition. RRO contributed to methodology, investigation, data curation, formal analysis, validation, visualization, and writing (review and editing). LR contributed to methodology, investigation, data curation, formal analysis, validation, and writing (original draft, review, and editing). HPS contributed to methodology, investigation, formal analysis, and writing (review and editing). HA contributed to conceptualization, investigation, writing (review and editing), and funding acquisition. CMM contributed to conceptualization, investigation, writing (review and editing), and funding acquisition. KM contributed to conceptualization, investigation, writing (review and editing), and funding acquisition.
Conflicts of Interest
Responsible DAI solutions Assessment Tool.XLSX File (Microsoft Excel File), 1305 KB
Methodological details and data sets.PDF File (Adobe PDF File), 1577 KB
- Sheikh A, Anderson M, Albala S, Casadei B, Franklin BD, Richards M, et al. Health information technology and digital innovation for national learning health and care systems. Lancet Digit Health 2021 Jun;3(6):e383-e396 [https://linkinghub.elsevier.com/retrieve/pii/S2589-7500(21)00005-4] [CrossRef] [Medline]
- Crawford K. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. London, UK: Yale University Press; 2021.
- Ethics guidelines for trustworthy AI. European Commission Independent High-Level Expert Group on Artificial Intelligence. 2019 Apr 08. URL: https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai [accessed 2022-11-30]
- Loh E. Medicine and the rise of the robots: a qualitative review of recent advances of artificial intelligence in health. BMJ Leader 2018 Jun 01;2(2):59-63 [https://bmjleader.bmj.com/content/leader/2/2/59.full.pdf] [CrossRef]
- Sharon T. When digital health meets digital capitalism, how many common goods are at stake? Big Data Soc 2018 Jul;5(2):205395171881903 [https://journals.sagepub.com/doi/10.1177/2053951718819032] [CrossRef]
- Obermeyer Z, Topol EJ. Artificial intelligence, bias, and patients' perspectives. Lancet 2021 May 29;397(10289):2038 [CrossRef] [Medline]
- The Lancet Digital Health. It's not easy being green. Lancet Digit Health 2021 Dec;3(12):e751 [https://linkinghub.elsevier.com/retrieve/pii/S2589-7500(21)00257-0] [CrossRef] [Medline]
- Parvez SM, Jahan F, Brune MN, Gorman JF, Rahman MJ, Carpenter D, et al. Health consequences of exposure to e-waste: an updated systematic review. Lancet Planet Health 2021 Dec;5(12):e905-e920 [https://linkinghub.elsevier.com/retrieve/pii/S2542-5196(21)00263-1] [CrossRef] [Medline]
- Sieck CJ, Sheon A, Ancker JS, Castek J, Callahan B, Siefer A. Digital inclusion as a social determinant of health. NPJ Digit Med 2021 Mar 17;4(1):52 [https://doi.org/10.1038/s41746-021-00413-8] [CrossRef] [Medline]
- The Lancet Digital Health. Digital technologies: a new determinant of health. Lancet Digit Health 2021 Nov;3(11):e684 [https://linkinghub.elsevier.com/retrieve/pii/S2589-7500(21)00238-7] [CrossRef] [Medline]
- Romanello M, McGushin A, Di Napoli C, Drummond P, Hughes N, Jamart L, et al. The 2021 report of the Lancet Countdown on health and climate change: code red for a healthy future. Lancet 2021 Oct 30;398(10311):1619-1662 [CrossRef] [Medline]
- van Wynsberghe A. Sustainable AI: AI for sustainability and the sustainability of AI. AI Ethics 2021 Feb 26;1(3):213-218 [https://link.springer.com/article/10.1007/s43681-021-00043-6] [CrossRef]
- Goirand M, Austin E, Clay-Williams R. Sustainable AI: AI implementing ethics in healthcare AI-based applications: a scoping review. Sci Eng Ethics 2021 Sep 03;27(5):61 [CrossRef] [Medline]
- Benrimoh D, Israel S, Perlman K, Fratila R, Krause M. Meticulous transparency—an evaluation process for an agile AI regulatory scheme. In: Proceedings of the 31st International Conference on Industrial Engineering and Other Applications of Applied Intelligent Systems on Recent Trends and Future Technology in Applied Intelligence. 2018 Presented at: IEA/AIE '18; June 25-28, 2018; Montreal, QC p. 869-880 URL: https://link.springer.com/chapter/10.1007/978-3-319-92058-0_83 [CrossRef]
- Nebeker C, Bartlett Ellis RJ, Torous J. Development of a decision-making checklist tool to support technology selection in digital health research. Transl Behav Med 2020 Oct 08;10(4):1004-1015 [https://europepmc.org/abstract/MED/31120511] [CrossRef] [Medline]
- Wiens J, Saria S, Sendak M, Ghassemi M, Liu VX, Doshi-Velez F, et al. Do no harm: a roadmap for responsible machine learning for health care. Nat Med 2019 Sep;25(9):1337-1340 [CrossRef] [Medline]
- Lehoux P, Rivard L, de Oliveira RR, Mörch CM, Alami H. Tools to foster responsibility in digital solutions that operate with or without artificial intelligence: a scoping review for health and innovation policymakers. Int J Med Inform 2023 Feb;170:104933 [https://linkinghub.elsevier.com/retrieve/pii/S1386-5056(22)00247-7] [CrossRef] [Medline]
- Floridi L. Translating principles into practices of digital ethics: five risks of being unethical. Philos Technol 2019 May 23;32(2):185-193 [https://link.springer.com/article/10.1007/s13347-019-00354-x] [CrossRef]
- Silva HP, Lehoux P, Miller FA, Denis JL. Introducing responsible innovation in health: a policy-oriented framework. Health Res Policy Syst 2018 Sep 10;16(1):90 [https://health-policy-systems.biomedcentral.com/articles/10.1186/s12961-018-0362-5] [CrossRef] [Medline]
- Abrishami P, Repping S. Nurturing societal values in and through health innovations comment on "what health system challenges should responsible innovation in health address?". Int J Health Policy Manag 2019 Oct 01;8(10):613-615 [https://europepmc.org/abstract/MED/31657189] [CrossRef] [Medline]
- Lehoux P, Silva HP, Oliveira RR, Rivard L. The responsible innovation in health tool and the need to reconcile formative and summative ends in RRI tools for business. J Responsible Innov 2020;7(3):646-671 [https://www.tandfonline.com/doi/full/10.1080/23299460.2020.1844974] [CrossRef]
- Silva HP, Lehoux P, Hagemeister N. Developing a tool to assess responsibility in health innovation: results from an international delphi study. Health Policy Technol 2018 Dec;7(4):388-396 [https://www.sciencedirect.com/science/article/abs/pii/S2211883718301485] [CrossRef]
- Silva HP, Lefebvre AA, Oliveira RR, Lehoux P. Fostering responsible innovation in health: an evidence-informed assessment tool for innovation stakeholders. Int J Health Policy Manag 2021 Mar 14;10(4):181-191 [https://europepmc.org/abstract/MED/32610749] [CrossRef] [Medline]
- Stilgoe J, Owen R, Macnaghten P. Developing a framework for responsible innovation. Res Policy 2013 Nov;42(9):1568-1580 [https://www.sciencedirect.com/science/article/pii/S0048733313000930] [CrossRef]
- Lehoux P, Rivard L, Silva HP. Responsible Innovation in Health: Concepts and Tools for Sustainable Impact. London, UK: Palgrave Macmillan; 2022.
- Zurynski Y, Herkes-Deane J, Holt J, McPherson E, Lamprell G, Dammery G, et al. How can the healthcare system deliver sustainable performance? A scoping review. BMJ Open 2022 May 24;12(5):e059207 [https://bmjopen.bmj.com/lookup/pmidlookup?view=long&pmid=35613812] [CrossRef] [Medline]
- Alami H, Rivard L, Lehoux P, Hoffman SJ, Cadeddu SB, Savoldelli M, et al. Artificial intelligence in health care: laying the Foundation for Responsible, sustainable, and inclusive innovation in low- and middle-income countries. Global Health 2020 Jun 24;16(1):52 [https://globalizationandhealth.biomedcentral.com/articles/10.1186/s12992-020-00584-1] [CrossRef] [Medline]
- Ricciardi F, Rossignoli C, Zardini A. Grand challenges and entrepreneurship: emerging issues, research streams, and theoretical landscape. Int Entrep Manag J 2021 Oct 07;17(4):1673-1705 [https://link.springer.com/article/10.1007/s11365-021-00771-5] [CrossRef]
- Gupta P, Srivastava R. Research on social enterprises from an emerging economy—systematic literature review and future research directions. J Soc Entrep 2021:1-36 [https://www.tandfonline.com/doi/abs/10.1080/19420676.2021.1974926?journalCode=rjse20] [CrossRef]
- Lee SD, Iott B, Banaszak-Holl J, Shih SF, Raj M, Johnson KE, et al. Application of mixed methods in health services management research: a systematic review. Med Care Res Rev 2022 Jun;79(3):331-344 [CrossRef] [Medline]
- Creswell JW, Clark VL. Designing and Conducting Mixed Methods Research. 3rd edition. Thousand Oaks, CA: SAGE Publications; 2017.
- Research Protocol. URL: https://udemontreal-my.sharepoint.com/:b:/g/personal/pascale_lehoux_umontreal_ca/EeXgNYq8-K1 PvD6pU9M1Uf8ByejYmrKusbxrj-Ob9W4Znw?e=1XHxvm [accessed 2022-11-30]
- Trochim WM, McLinden D. Introduction to a special issue on concept mapping. Eval Program Plann 2017 Feb;60:166-175 [https://doi.org/10.1016/j.evalprogplan.2016.10.006] [CrossRef] [Medline]
- Campbell KA, Orr E, Durepos P, Nguyen L, Li L, Whitmore C, et al. Reflexive thematic analysis for applied qualitative health research. Qual Rep 2021;26(6):2011-2028 [https://nsuworks.nova.edu/tqr/vol26/iss6/24/] [CrossRef]
- Lozano S, Calzada-Infante L, Adenso-Díaz B, García S. Complex network analysis of keywords co-occurrence in the recent efficiency analysis literature. Scientometrics 2019 Jun 7;120(2):609-629 [https://link.springer.com/article/10.1007/s11192-019-03132-w] [CrossRef]
- Gwet KL. Handbook of Inter-Rater Reliability: The Definitive Guide to Measuring the Extent of Agreement Among Raters. Oxford, MS: Advanced Analytics; 2014.
- Morley J, Floridi L, Kinsey L, Elhalal A. From what to how: an initial review of publicly available ai ethics tools, methods and research to translate principles into practices. Sci Eng Ethics 2020 Aug;26(4):2141-2168 [https://europepmc.org/abstract/MED/31828533] [CrossRef] [Medline]
- Torous J, Stern AD, Bourgeois FT. Regulatory considerations to keep pace with innovation in digital health products. NPJ Digit Med 2022 Aug 19;5(1):121 [https://doi.org/10.1038/s41746-022-00668-9] [CrossRef] [Medline]
- Reddy S, Allan S, Coghlan S, Cooper P. A governance model for the application of AI in health care. J Am Med Inform Assoc 2020 Mar 01;27(3):491-497 [https://europepmc.org/abstract/MED/31682262] [CrossRef] [Medline]
- Siala H, Wang Y. SHIFTing artificial intelligence to be responsible in healthcare: a systematic review. Soc Sci Med 2022 Mar;296:114782 [https://linkinghub.elsevier.com/retrieve/pii/S0277-9536(22)00085-5] [CrossRef] [Medline]
- Classification of digital health interventions v1.0: a shared language to describe the uses of digital technology for health. World Health Organization. 2018. URL: https://apps.who.int/iris/bitstream/handle/10665/260480/WHO-RHR-18.06-eng.pdf [accessed 2022-11-30]
- Areas for future action in the responsible AI ecosystem. The Future Society. 2020. URL: https://gpai.ai/projects/responsible-ai/areas-for-future-action-in-responsible-ai.pdf [accessed 2022-11-30]
- Benjamens S, Dhunnoo P, Meskó B. The state of artificial intelligence-based FDA-approved medical devices and algorithms: an online database. NPJ Digit Med 2020 Sep 11;3(1):118 [https://doi.org/10.1038/s41746-020-00324-0] [CrossRef] [Medline]
- Muehlematter UJ, Daniore P, Vokinger KN. Approval of artificial intelligence and machine learning-based medical devices in the USA and Europe (2015–20): a comparative analysis. Lancet Digit Health 2021 Mar;3(3):e195-e203 [CrossRef]
- Fjeld J, Achten N, Hilligoss H, Nagy A, Srikumar M. Principled artificial intelligence: mapping consensus in ethical and rights-based approaches to principles for AI. SSRN Journal Preprint posted online February 14, 2020 [https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3518482] [CrossRef]
- Hagendorff T. The ethics of AI ethics: an evaluation of guidelines. Minds Mach 2020 Feb 01;30(1):99-120 [https://link.springer.com/article/10.1007/s11023-020-09517-8] [CrossRef]
- CADTH Horizon Scanning Service. List of 2021 health technology trends to watch. Can J Health Technol 2021 Jan 29;1(1) [https://www.sciencegate.app/document/10.51731/cjht..29] [CrossRef]
- AI for sustainable development goals (AI4SDGs) think tank. Center for Long-term Artificial Intelligence. 2020. URL: https://ai-for-sdgs.academy/ [accessed 2022-11-30]
- Terwee CB, Prinsen CA, Chiarotto A, Westerman MJ, Patrick DL, Alonso J, et al. COSMIN methodology for evaluating the content validity of patient-reported outcome measures: a Delphi study. Qual Life Res 2018 May;27(5):1159-1170 [https://europepmc.org/abstract/MED/29550964] [CrossRef] [Medline]
- Lam K, Abràmoff MD, Balibrea JM, Bishop SM, Brady RR, Callcut RA, et al. A Delphi consensus statement for digital surgery. NPJ Digit Med 2022 Jul 19;5(1):100 [https://doi.org/10.1038/s41746-022-00641-6] [CrossRef] [Medline]
- Landis JR, Koch GG. The measurement of observer agreement for categorical data. Biometrics 1977 Mar;33(1):159-174 [Medline]
- Tricco AC, Lillie E, Zarin W, O'Brien KK, Colquhoun H, Levac D, et al. PRISMA extension for Scoping Reviews (PRISMA-ScR): checklist and explanation. Ann Intern Med 2018 Oct 02;169(7):467-473 [https://www.acpjournals.org/doi/abs/10.7326/M18-0850?url_ver=Z39.88-2003&rfr_id=ori:rid:crossref.org&rfr_dat=cr_pub 0pubmed] [CrossRef] [Medline]
- Itten R, Hischier R, Andrae AS, Bieser JC, Cabernard L, Falke A, et al. Digital transformation—life cycle assessment of digital services, multifunctional devices and cloud computing. Int J Life Cycle Assess 2020 Aug 12;25(10):2093-2098 [https://link.springer.com/article/10.1007/s11367-020-01801-0] [CrossRef]
- Clarke R. Principles and business processes for responsible AI. Comput Law Secur Rev 2018 Aug;35(4):410-422 [https://doi.org/10.1016/j.clsr.2019.04.007] [CrossRef]
- Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science 2019 Oct 25;366(6464):447-453 [CrossRef] [Medline]
- Hernandez-Boussard T, Bozkurt S, Ioannidis JP, Shah NH. MINIMAR (MINimum Information for Medical AI Reporting): developing reporting standards for artificial intelligence in health care. J Am Med Inform Assoc 2020 Dec 09;27(12):2011-2015 [https://europepmc.org/abstract/MED/32594179] [CrossRef] [Medline]
- Noel-Brune M, Goldizen FC, Neira M, van den Berg M, Lewis N, King M, et al. Health effects of exposure to e-waste. Lancet Glob Health 2013 Aug;1(2):e70 [https://linkinghub.elsevier.com/retrieve/pii/S2214-109X(13)70020-2] [CrossRef] [Medline]
- Tiffin N, George A, LeFevre AE. How to use relevant data for maximal benefit with minimal risk: digital health data governance to protect vulnerable populations in low-income and middle-income countries. BMJ Glob Health 2019 Apr 11;4(2):e001395 [https://gh.bmj.com/lookup/pmidlookup?view=long&pmid=31139457] [CrossRef] [Medline]
- Falco G, Shneiderman B, Badger J, Carrier R, Dahbura A, Danks D, et al. Governing AI safety through independent audits. Nat Mach Intell 2021 Jul 20;3(7):566-571 [https://uwe-repository.worktribe.com/output/7562797/governing-ai-safety-through-independent-audits] [CrossRef]
- Sarkar S, Mateus S. Doing more with less - how frugal innovations can contribute to improving healthcare systems. Soc Sci Med 2022 Aug;306:115127 [https://linkinghub.elsevier.com/retrieve/pii/S0277-9536(22)00433-6] [CrossRef] [Medline]
- Bhatti Y, Taylor A, Harris M, Wadge H, Escobar E, Prime M, et al. Global lessons in frugal innovation to improve health care delivery in the united states. Health Aff (Millwood) 2017 Nov;36(11):1912-1919 [http://hdl.handle.net/10044/1/51814] [CrossRef] [Medline]
- Dabic M, Obradovic T, Vlacic B, Sahasranamam S, Paul J. Frugal innovations: A multidisciplinary review and agenda for future research. J Bus Res 2022 Mar;142:914-929 [https://doi.org/10.1016/j.jbusres.2022.01.032] [CrossRef]
- Weyrauch T, Herstatt C. What is frugal innovation? Three defining criteria. J Frugal Innov 2016 Dec 27;2(1) [CrossRef]
- Lokmic-Tomkins Z, Davies S, Block LJ, Cochrane L, Dorin A, von Gerich H, et al. Assessing the carbon footprint of digital health interventions: a scoping review. J Am Med Inform Assoc 2022 Nov 14;29(12):2128-2139 [https://europepmc.org/abstract/MED/36314391] [CrossRef] [Medline]
- Moultrie J, Sutcliffe L, Maier A. Exploratory study of the state of environmentally conscious design in the medical device industry. J Clean Prod 2015 Dec;108:363-376 [https://www.sciencedirect.com/science/article/pii/S0959652615007349] [CrossRef]
- Home page. In Fieri. URL: https://infieri.online/en/home/ [accessed 2022-11-30]
|AI: artificial intelligence|
|DAI solutions: digital health solutions that operate with or without artificial intelligence|
|PRISMA-ScR: Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews|
|RIH: Responsible Innovation in Health|
Edited by A Mavragani; submitted 25.04.23; peer-reviewed by P Cooper, P Mechael; comments to author 17.06.23; revised version received 27.06.23; accepted 26.07.23; published 28.08.23Copyright
©Pascale Lehoux, Robson Rocha de Oliveira, Lysanne Rivard, Hudson Pacifico Silva, Hassane Alami, Carl Maria Mörch, Kathy Malas. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 28.08.2023.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included.