Barriers to and facilitators of artificial intelligence adoption in health care: scoping review

Probably not a substantial claim: may not pass S3 (load-bearing).

An estimated 70% of health IT projects fail.

“An estimated 70% of health IT projects fail [22]”

• cites [22] sourcePartly supported

  The Abstract says HIT projects fail at a rate *up to* 70%, using a definition of failure that includes delays, substantial cost overruns, unmet goals, and abandonment. The citing sentence presents 70% as the estimated failure rate rather than an upper bound, and omits that definition. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The scoping review was guided by the Joanna Briggs Institute methodology and the Arksey and O'Malley framework.

“This review was guided by the methodology and reporting structure outlined for scoping reviews by the Joanna Briggs Institute as well as Arksey and O’Malley [28].”

• cites [28] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Morrison quoted stakeholders seeking clarity on the extent to which there is a need to provide transparency on AI output to patients and how this is directed by legislation or data protection laws.

“Morrison [36] particularly quoted stakeholders seeking clarity on the extent to which there is a need to provide transparency on AI output to patients and how this is directed by legislation or data protection laws.”

• cites [36] sourceSupported

Classified as probably not a substantial claim.

Improving algorithmic transparency will be a key consideration to change attitudes and build the trust of adopters.

“To facilitate adoption, improving algorithmic transparency will be a key consideration to change attitudes and build the trust of adopters [35,40].”

• cites [35] sourceSupported
• cites [40] sourceSupported

May combine two claims.

Baxter et al reported concerns from adopters around the lack of explainability regarding the prediction of the AI algorithm embedded in the EHR to predict unplanned readmission; specifically, the lack of explainability regarding what features of the algorithm were driving the output was an impediment to trust among adopters.

“Baxter et al [35] reported concerns from adopters around the lack of explainability regarding the prediction of the AI algorithm embedded in the EHR to predict unplanned readmission; specifically, the lack of explainability regarding what features of the algorithm were driving the output was an impediment to trust among adopters.”

• cites [35] sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Successful adoption was achieved by publishing evidence on the algorithm's underpinnings and providing clinicians with details on how data elements interacted within the algorithm to produce the predictive output.

“Moorman [38] reported that successful adoption was achieved by publishing evidence on the algorithm’s underpinnings and providing clinicians with details on how data elements interacted within the algorithm to produce the predictive output.”

• cites [38] sourcePartly supported

  The Framework section says the authors presented an existing published evidence base to clinicians and used patient-focused sessions to show how data elements produced a risk score, helping develop confidence in the tool. The Conclusion identifies understandable science as an ingredient of successful implementation. The source does not say that publishing evidence was the adoption strategy; it says the authors presented evidence that had already been published. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

Investing into explainable AI that produces a transparent and understandable AI could help address the issue of acceptability.

“Weinert et al [45] recommended that investing into explainable AI that produces a transparent and understandable AI could help address the issue of acceptability.”

• cites [45] sourceSupported

May combine two claims.

Inaccurate and underrepresentative training data sets for AI models can cause bias, misleading predictions, adverse events, and large-scale discrimination, causing barriers to adoption.

“Inaccurate and underrepresentative training data sets for AI models can cause bias, misleading predictions, adverse events, and large-scale discrimination, causing barriers to adoption [32,41,47,48].”

• cites [32] sourcePartly supported

  The Main body section on barriers says unreliable training data obstructs implementation and that dataset bias can make AI predictions less accurate. It also describes unrepresentative datasets and risks of bias affecting marginalized groups. Our copy does not establish adverse events or large-scale discrimination as consequences of those datasets. (LLM, conf=0.87)

  Experimental — accuracy not yet measured
• cites [41] sourcePartly supported

  In “Equity of access” (page unavailable), the paper says under-represented groups may receive inaccurate diagnoses and warns that algorithmic bias can exacerbate inequities. “Challenges in clinical deployment of AI” (page unavailable) says training datasets and validation require consideration before adoption. The copy does not specifically establish that inaccurate training data cause large-scale discrimination or adverse events. (LLM, conf=0.80)

  Experimental — accuracy not yet measured
• cites [47] sourcePartly supported

  The Equity section (page unavailable) explicitly says inaccurate and underrepresentative training data can cause bias, misleading predictions, adverse events, and large-scale discrimination. The Challenges and benefits section identifies data bias as an application problem and early adoption as a challenge, but does not explicitly say those data-related harms cause barriers to adoption. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [48] sourcePartly supported

  The Discussion (Principal Findings; page not available) identifies concerns about systemic bias in data used by ML tools and patient safety as factors affecting confidence in implementation. The Conclusions also reports concerns about accuracy and reliability. The paper does not establish that inaccurate or underrepresentative training sets cause misleading predictions, adverse events, or large-scale discrimination; it reports perceived barriers rather than those outcomes. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

May combine two claims.

Inadequate data from representative groups, algorithms designed to represent a majority, and missing variables that impact predictions are components that contribute to bias.

“Inadequate data from representative groups, algorithms designed to represent a majority, and missing variables that impact predictions are components that contribute to bias [21].”

• cites [21] sourceSupported

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing).

AI systems should support the understanding and explanation of the causal models as opposed to simply solving through pattern recognition.

“In health care, causality is especially important when using automated decision systems; therefore, Holzinger et al [33] emphasized that AI systems should support the understanding and explanation of the causal models as opposed to simply solving through pattern recognition.”

• cites [33] sourceSupported

Classified as probably not a substantial claim. May combine two claims.

Revealing the process of how the algorithm was developed, who was involved in the development process, whether clinicians were consulted, and how the data were processed would enable acceptability.

“Furthermore, revealing the process of how the algorithm was developed, who was involved in the development process, whether clinicians were consulted, and how the data were processed would enable acceptability [42,43].”

• cites [42] sourcePartly supported

  The Results section ('Views towards using mHealth apps for skin cancer screening'; page not provided) identifies transparency about the app-development team and whether an algorithm or clinician performs the assessment as facilitators of trust and use. It does not say that explaining how the algorithm was developed, whether clinicians were consulted during its development, or how the app processes data would increase acceptability. The Methods section's discussion of data analysis concerns the researchers' focus-group transcripts, not the app's data processing. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [43] sourcePartly supported

  The Results section, “Ethical and Lawful Processing of Patient Data by AI Applications” (page unavailable), reports agreement on the need for transparency about data processing. It does not say that disclosing how an algorithm was developed, who developed it, or whether clinicians were consulted would increase acceptability. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

May combine two claims.

In some studies, patients reported the need to assess the usefulness of the AI before using it and had concerns around the quality and accuracy of the output, thereby questioning the value of AI as a whole.

“In some studies, patients reported the need to assess the usefulness of the AI before using it and had concerns around the quality and accuracy of the output, thereby questioning the value of AI as a whole [39,42,52-55].”

• cites [39] sourcePartly supported

  The Results (AI hesitancy; page unavailable) report participants’ concerns about the quality and accuracy of chatbot information and hesitation about using chatbots in healthcare. The Results (Quantitative analysis; page unavailable) link perceived usefulness to acceptability. The paper does not report that patients said they needed to assess usefulness before use, nor that they questioned the value of AI as a whole rather than particular healthcare chatbot uses. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [42] sourcePartly supported

  In Results (‘Views towards using mHealth apps for skin cancer screening’; page not specified), a participant says they would research how well the algorithm works, and the study reports concerns about screening accuracy and the apps’ value. The study concerns people in the Dutch general population considering skin-cancer screening apps; it does not establish that patients questioned the value of AI as a whole. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [52] sourcePartly supported

  The Discussion (Principal Findings; page unavailable) says people may assess an ophthalmic AI device’s usefulness before switching from a face-to-face eye examination. The Methods (Measurement; page unavailable) includes survey items about inaccurate diagnoses and incorrect screening results. But the paper does not report that these concerns led respondents to question the value of AI as a whole; the Discussion says perceived risks did not affect intention to use the devices. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [53] sourcePartly supported

  Section 4.6 (p. 24) reports that elderly participants and their families distrusted AI-enabled wearable medical devices and questioned their clinical efficiency; family members also doubted whether advertised AI features provided useful services. Section 4.3 (p. 20) raises the possibility of misleading analytic results, but in relation to clinicians. The paper does not report patients saying they needed to assess AI's usefulness before use or questioning the value of AI as a whole. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [54] sourceNot checked — no full text of the source
• cites [55] sourcePartly supported

  The Discussion’s Analysis of Results (page unavailable) reports that perceived usefulness influenced patients’ motivation to adopt AI-based home care systems. The Methods’ Accountability and Security section mentions concerns about reliability, but the paper does not report patients assessing output quality or accuracy before use, or questioning the value of AI as a whole. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

May combine two claims.

Ease of use; complex interfaces; and inconsistent performance, for example, due to false positives and negatives, add burden to the workflow, creating more work for clinicians, thereby impeding adoption.

“Ease of use; complex interfaces; and inconsistent performance, for example, due to false positives and negatives add burden to the workflow, creating more work for clinicians, thereby impeding adoption [42,52,57-59].”

• cites [42] sourcePartly supported

  The Results (page not provided) identify complex interfaces and perceived lack of screening accuracy as barriers, and ease of use as a facilitator. Participants raised concerns about incorrect assessments, but the study does not show that app performance adds to clinicians’ workflow or creates more work for them; it reports views from the general population. (LLM, conf=0.87)

  Experimental — accuracy not yet measured
• cites [52] sourcePartly supported

  The Methods > Measurement section considers ease of use, possible performance failure, and incorrect screening results for ophthalmic AI devices. But the study examines the public’s intention to use those devices, not added work for clinicians or complex interfaces; the Discussion > Principal Findings reports no significant direct effect of ease of use on intention to use. The cited paper therefore supports some concerns named in the sentence, but not its clinician-workflow explanation for impeded adoption. Page numbers are not available in our copy. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [57] sourceSupported
• cites [58] sourcePartly supported

  The Results sections on Ease of use and The importance of trust describe minor usability problems and say that trust in diagnostic guidance matters for adoption into the clinical workflow. The Discussion also links ease of use to time efficiency. The paper does not establish that complex interfaces or false positives and negatives created extra work for clinicians; respondents generally described this CDSS as easy to use and time-saving. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [59] sourcePartly supported

  The Results (Technology Adoption Considerations; page unavailable) report that physicians wanted an easy-to-use platform and might revert to old practices if it was too troublesome to use. The Discussion (Principal Findings; page unavailable) likewise links usability to adoption. Our copy does not establish that false positives, false negatives, or inconsistent performance create additional clinical work and thereby impede adoption. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

Dissatisfaction was reported if recommendations were too similar, inappropriate, or not useful.

“This also included dissatisfaction if recommendations were too similar, inappropriate, or not useful [35,37,38,40,56].”

• cites [35] sourcePartly supported

  In Results > Concerns Regarding Model Implementation (page unavailable), stakeholders questioned whether the risk score was useful and whether its predictors were relevant and actionable. This supports concerns about usefulness and appropriateness, but the paper discusses risk scores rather than recommendations and does not report dissatisfaction because recommendations were too similar. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [37] sourceCould not verify

  The Results discuss physicians’ concerns about the utility of AI tools, and the Discussion mentions limited usefulness of point-of-care tests. Neither passage reports dissatisfaction with recommendations because they were too similar, inappropriate, or not useful. Searches for those terms and paraphrases found no passage addressing that claim. (LLM, conf=0.73)

  Experimental — accuracy not yet measured
• cites [38] sourceCould not verify

  The Framework and Principles for Implementing Predictive Analytics Monitoring sections discuss clinicians’ skepticism about a predictive score and how to make the tool relevant to their work. They do not report dissatisfaction because recommendations were too similar, inappropriate, or not useful (searched for those terms and related descriptions of dissatisfaction). No page numbers were available in our copy. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [40] sourcePartly supported

  The Clinical utility section reports informants’ concern that some models lacked clear actionability and that poorly configured alerts reduced clinical utility. This supports concern about usefulness, but the paper does not report dissatisfaction with recommendations because they were too similar or inappropriate. (LLM, conf=0.79)

  Experimental — accuracy not yet measured
• cites [56] sourcePartly supported

  In Results > Qualitative focus group sessions/interviews (page not provided), a cardiologist says information that is not actionable or helpful slows clinicians down and is ignored. The paper also reports concern about AI making wrong clinical decisions, but it does not report dissatisfaction with recommendations because they were too similar, or specifically describe inappropriate recommendations. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

If an algorithm is not performing up to standards or adds more work to the clinicians, this can impact adoption, as clinicians perceive it as a high risk.

“Finally, Choudhury [61] noted that if an algorithm is not performing up to standards or adds more work to the clinicians, this can impact adoption, as clinicians perceive it as a high risk.”

• cites [61] sourcePartly supported

  The Discussion (page unavailable) reports that poorer effort and performance expectancy worsened perceived risk, and that greater perceived risk reduced clinicians’ intention to use the Blood Utilization Calculator. The Instrumentation section defines these expectancies as ease of use and perceived usefulness, but the paper does not specifically establish that an algorithm adding to clinicians’ workload causes them to perceive it as high risk. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Risk of harm can be lowered if AI algorithms are developed with the 5 rights, similar to other clinical decision support tools.

“The risk of harm can be lowered if AI algorithms are developed with the 5 rights (in the case of an automated decision system), similar to other clinical decision support tools [40].”

• cites [40] sourcePartly supported

  In “Recommendations for institutions developing clinical PM and ML applications” (page not provided), the paper says PM and ML models should be developed like other clinical decision-support tools and address the five rights. It describes this approach as useful, but does not say that following the five rights lowers the risk of harm, including for automated decision systems. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

May combine two claims.

Lack of accuracy, doubts about unsafe results, privacy breaches, patients' perceptions and acceptance of automation, and uncertainty about developers' reliability, availability, usability, and perceived usefulness were found to be obstacles to gaining trust.

“Lack of accuracy; doubts about unsafe results; privacy breaches; patients’ perceptions and acceptance of automation; and uncertainty about developers’ reliability, availability, usability, and perceived usefulness were found to be obstacles to gaining trust [39,42,47,50,53,54,69,70].”

• cites [39] sourcePartly supported

  The Results (AI hesitancy) report concerns about inaccurate or harmful chatbot advice and breaches of confidentiality. The Discussion links accuracy, privacy and trustworthiness concerns to reduced adoption, and the survey relates perceived utility and attitudes to acceptability. The paper does not identify developers’ reliability, availability or usability as obstacles to gaining trust; its findings chiefly concern trust in and acceptance of chatbots. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [42] sourcePartly supported

  The Results (page unavailable) identify perceived inaccuracy and doubts about developers’ reliability as sources of untrustworthiness. They also identify concerns about incorrect assessments, privacy, preference for a doctor over an algorithm, limited perceived value, and a complex interface as barriers to using skin-cancer screening apps. The paper does not identify availability as a barrier, and it does not classify every listed barrier specifically as an obstacle to gaining trust. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [47] sourcePartly supported

  The Equity section links potentially unsafe AI results to an obstacle for doctors accepting AI, and the Introduction and Challenges and benefits sections discuss system errors and privacy risks. The paper also identifies lack of consideration for users’ perspectives as an adoption challenge, but these passages do not establish patients’ acceptance of automation or uncertainty about developers’ reliability, availability, usability, and perceived usefulness as obstacles to trust. Page numbers are unavailable in our copy. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [50] sourceNot checked — no full text of the source
• cites [53] sourcePartly supported

  The Further discussion (pp. 25, 27) links device-performance concerns, privacy breaches, psychological resistance, and doubts about clinical evidence to lack of trust in wearable medical devices and their providers. The Related studies section (p. 7) identifies perceived usefulness and ease of use as factors in acceptance, but does not establish them—or developers’ availability and usability—as obstacles to gaining trust. The paper discusses unsafe outcomes as a concern, but not every item in the citing sentence is supported as a trust barrier. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [54] sourceNot checked — no full text of the source
• cites [69] sourcePartly supported

  The Results (‘Patients’ views on AI in radiology,’ page not specified) report moderate distrust concerning diagnostic accuracy, communication, and confidentiality; the questionnaire also records worries about scans analyzed without human involvement and personal data falling into the wrong hands. These support concerns about accuracy, privacy, and acceptance of automation, but the paper does not report actual privacy breaches or identify developers’ reliability, availability, usability, and perceived usefulness as obstacles to trust. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [70] sourcePartly supported

  Results §3.6 (page unavailable) reports trust concerns about AI accuracy, possible harm from malfunctions, patient acceptance, and developers’ lack of clinical knowledge. It also identifies availability and patients’ perceptions as barriers. The passages do not identify privacy breaches or perceived usefulness as obstacles to trust, and do not clearly frame usability that way. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

May combine two claims.

Skepticism and a lack of understanding were seen as barriers to AI adoption.

“Skepticism and a lack of understanding were seen as barriers to AI adoption [48,51].”

• cites [48] sourceSupported
• cites [51] sourceSupported

May combine two claims.

Facilitators of trust include endorsement of the technology by experts and academically backed clinicians (including regulating bodies such as the government), evidence of output accuracy based on evaluation of AI, and positive opinions from trusted thought leaders in clinical fields.

“Facilitators of trust included the endorsement of the technology by experts as well as academically backed clinicians, including regulating bodies such as the government; evidence of output accuracy based on the evaluation of AI; and positive opinions from trusted thought leaders in the respective clinical fields [42,58,60].”

• cites [42] sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [58] sourcePartly supported

  The Results section, “Factors governing trust” (page unavailable), identifies clinical evidence of diagnostic accuracy, certification, and recommendations or positive opinions from dermatologists as factors that could increase trust. It does not identify government bodies or academically backed clinicians as endorsers, so that part of the claim is not supported by this source. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [60] sourcePartly supported

  Section 4.3, “Trust and Credibility of ChatGPT” (page not provided), says that evidence-based data on ChatGPT’s accuracy and reliability could enhance healthcare workers’ trust. The cited paper does not identify endorsement by experts, academically backed clinicians, or government bodies, or positive opinions from clinical thought leaders, as facilitators of trust. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

Shortages of personnel with the required skills and the quality of IT infrastructure available for AI implementation were reported as barriers.

“Shortages of personnel with the required skills were reported as barriers, along with the quality of IT infrastructure available for AI implementation [32,35,36,40,45,50].”

• cites [32] sourcePartly supported

  The ‘Human resources’ and ‘Education’ subsections under ‘Barriers involved in AI implementation’ identify the need for knowledgeable personnel and training, supporting the skilled-personnel portion. Our copy does not identify the quality of available IT infrastructure as a barrier; searches for infrastructure and related technical resources found no such passage. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [35] sourceCould not verify

  The Results (Concerns Regarding Model Implementation; page not provided) discuss staff-training time, and the Conclusion (page not provided) lists workflow and adoption barriers. Searches for skilled-personnel shortages and IT-infrastructure quality found no passage reporting either as a barrier. (LLM, conf=0.72)

  Experimental — accuracy not yet measured
• cites [36] sourcePartly supported

  In “Socio-political context of the NHS in 2020” (page unavailable), nine participants mentioned the quality of NHS IT infrastructure as a barrier to AI adoption, though the Conclusion notes that participants disagreed. The paper discusses a need for education and for IT project managers, but the passages found do not report a shortage of personnel with the required skills as a barrier. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [40] sourceSupported
• cites [45] sourceSupported
• cites [50] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Radiographers and radiologists held more positive attitudes toward the adoption of AI because it would help address workforce shortages in the radiology field in the United Kingdom.

“In a different perspective on the shortage of resources, Chen et al [68] noted that radiographers and radiologists held more positive attitudes toward the adoption of AI, as it would help address workforce shortages in the radiology field in the United Kingdom.”

• cites [68] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Lack of technological infrastructure maturity to allow for the integration of AI was noted.

“Hickman et al [71], in particular, noted the lack of technological (infrastructure) maturity to allow for the integration of AI.”

• cites [71] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Lack of clinical and economic measurements is a barrier to practical implementation.

“Wolff et al [78] particularly noted the lack of clinical and economic measurements as a barrier to practical implementation.”

• cites [78] sourceSupported

May combine two claims.

Data security was identified as a major contributor toward hesitancy, with concerns around cybersecurity relating to both training and testing data as well as fear of trackers and spyware obtaining unsolicited data.

“Data security was identified as a major contributor toward hesitancy, with concerns around cybersecurity relating to both training and testing data as well as fear of trackers and spyware obtaining unsolicited data [39,41,42,45,48,77].”

• cites [39] sourcePartly supported

  The Discussion identifies cyber-security as a potential factor in AI hesitancy, and Results > AI hesitancy reports concerns about hacking or leakage of sensitive health information. The paper does not connect those concerns to training and testing data or mention trackers or spyware obtaining unsolicited data. Page numbers were not available in our copy. (LLM, conf=0.89)

  Experimental — accuracy not yet measured
• cites [41] sourcePartly supported

  In “Challenges in clinical deployment of AI” (page not provided), the paper identifies cybersecurity and data protection as issues for both training and testing data. The telemedicine-deployment section also lists security and privacy concerns among barriers to EHR adoption. It does not identify security as a major contributor to hesitancy or mention trackers or spyware obtaining unsolicited data. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [42] sourcePartly supported

  The Results section identifies privacy concerns as a barrier to using skin-cancer screening apps and says fear of unsolicited data sharing by trackers and spyware deterred participants. It does not report cybersecurity concerns about training or testing data, so that part of the claim is unsupported by this source. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [45] sourcePartly supported

  In Results, Table 2, 22 of 40 respondents (55%) identified data protection as a barrier to AI implementation and use. The paper does not link that barrier to cybersecurity of training or testing data, or to trackers and spyware obtaining unsolicited data. No page number is available in our copy. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [48] sourcePartly supported

  The Results (Social Media Provider and Nonprovider Survey; page not provided) identify privacy/security as a concern, and the Discussion (Principal Findings; page not provided) says it is a factor that needs addressing to build confidence. The paper does not specify cybersecurity risks involving training or testing data, or fears of trackers and spyware obtaining unsolicited data. Its survey results also do not establish data security as a major contributor to hesitancy. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [77] sourcePartly supported

  The Methodology’s ‘Multilayer recursive structural analysis’ section identifies data-leakage risk, including theft of sensitive medical data by hackers, as an important factor in medical institutions’ adoption decisions (page not available). It does not discuss cybersecurity of training or testing data, or trackers and spyware obtaining unsolicited data. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

May combine two claims.

Privacy and ethics concerns were focused on maintaining confidentiality, ensuring processes are in place to obtain consent, and having informed consent with clarity on how data will be processed.

“Concerns about privacy and ethics were focused on maintaining confidentiality, ensuring processes are in place to obtain consent, and having informed consent with clarity on how the data will be processed [32,39,47,56].”

• cites [32] sourcePartly supported

  The “Privacy and confidentiality” subsection identifies confidentiality and ethical data management as concerns. The following passage says patient consent is important and may need to be a dynamic process, but it does not say that informed consent must clearly explain how data will be processed. No page numbers are available in our copy. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [39] sourcePartly supported

  The Results (AI hesitancy and Qualitative analysis; page unavailable) report participants’ concerns about chatbot confidentiality, hacking, and data leakage. The Method describes consent obtained from study participants, not a concern about consent processes for chatbot users. The passages do not support the claim that concerns focused on informed consent or clarity about how chatbot data would be processed. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [47] sourcePartly supported

  The “2) Personal” section calls for patient confidentiality and adequate processes for obtaining consent. The copy does not specify that informed consent must clarify how data will be processed; the transparency discussion concerns explaining clinical decisions, not data processing. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [56] sourcePartly supported

  The Results (Qualitative focus group sessions/interviews; page unavailable) report concern about protecting patient data privacy, and the quantitative survey reports common HIPAA concerns. The paper does not describe concerns about processes for obtaining consent or explaining to patients how their data will be processed; its reported evaluation processes concern adoption of AI technology, not consent. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

The German government introduced a new law to help organizations bridge funding gaps for AI implementations.

“Weinert et al [45] noted that to overcome the barrier of lack of resources and meet the financial investment demands of AI implementations, the German government introduced a new law that could help organizations bridge funding gaps;”

• cites [45] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Financial barriers in the context of cost and benefits, the lack of a sustainable business model, and insufficient funding to meet public demands should also be considered.

“Sun and Medaglia [50] and Xing et al [53] reported that financial barriers in the context of cost and benefits, the lack of a sustainable business model, and insufficient funding to meet public demands should also be considered.”

• cites [50] sourceNot checked — no full text of the source
• cites [53] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Explanations are crucial to involving a human in the process of verifying the decision of the algorithm, for example, by revealing what features were used in training the AI algorithm.

“For example, Markus et al [81] noted that explanations are crucial to involving a human in the process of verifying the decision of the algorithm, for example, by revealing what features were used in training the AI algorithm.”

• cites [81] sourcePartly supported

  Section 3 (p. 8) says explanations can allow a human in the loop to detect and correct problems, and that clinicians want to know which features a model used to predict an outcome. The claim overstates this as crucial: the paper says explanations help only in some situations and are not necessary to satisfy other model goals. It also discusses features used for prediction, not specifically features used in training. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Convolutional neural networks will frequently underdiagnose Hispanic patients at a disproportionate rate due to the potential lack of access to health care and insurance type.

“Seyyed-Kalantari et al [87] noted that convolutional neural networks will frequently underdiagnose Hispanic patients at a disproportionate rate due to the potential lack of access to health care and insurance type.”

• cites [87] sourcePartly supported

  Section 7.1 (p. 8) reports that Hispanic patients were the least-favored racial subgroup for true-positive rates in the CXR dataset, supporting disproportionate missed diagnoses in that setting. Section 7.4 (p. 9) suggests that disadvantaged groups may use healthcare services less, but the paper does not establish that healthcare access or insurance type caused the Hispanic disparity; it analyzes insurance disparities separately. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Federated learning allows models to be trained on large data sets without sharing sensitive information.

“This technique allows for models to be trained on large data sets without sharing sensitive information.”

• cites [89] sourceSupported

May combine two claims.

Bias in AI systems can be introduced due to biased data, algorithms being trained on the biased data, limitations in the model itself, small training size, lack of user participation, and other unseen factors.

“Bias in AI systems can be introduced due to biased data, algorithms being trained on the biased data, limitations in the model itself, small training size, lack of user participation, and other unseen factors [84].”

• cites [84] sourceSupported

Classified as probably not a substantial claim. May combine two claims.

Challenges in addressing algorithmic bias include a lack of clear definitions and standards of 'fairness,' insufficient contextual specificity, and the 'black-box' nature of algorithms.

“According to Panch et al [46], several challenges need to be addressed when addressing algorithmic bias. They include a lack of clear definitions and standards of “fairness,” insufficient contextual specificity, and the “black-box” nature of algorithms.”

• cites [46] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Rogers et al shared a very practical approach to how they evaluated an AI model for ethics.

“Similarly, Rogers et al [97] shared a very practical approach to how they evaluated an AI model for ethics.”

• cites [97] sourceSupported

May combine two claims.

An algorithm widely used in US hospitals to identify patients who need additional care was found to use the cost expenditure by patients as a means to identify those who need extra care, which was discriminatory toward Black patients because they generally spend less than White patients on health care, resulting in false conclusions.

“Similarly, an algorithm that is widely used in US hospitals to identify patients who need additional care was found to use the cost expenditure by patients as a means to identify those who need extra care. This was discriminatory toward Black patients, as they generally spend less than White patients on health care, resulting in false conclusions [95].”

• cites [95] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Liao provides an example of a prediction of a 70% chance for a supposed patient's tumor to become malignant in 5 years; however, the algorithm does not necessarily provide detailed reasons as to how it arrived at the conclusion.

“They provide an example of a prediction of a 70% chance for a supposed patient’s tumor to become malignant in 5 years; however, the algorithm does not necessarily provide detailed reasons as to how it arrived at the conclusion.”

• cites [94] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

According to Liao, there are >80 ethical frameworks that have been proposed for AI.

“According to Liao [94], there are >80 ethical frameworks that have been proposed for AI.”

• cites [94] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

A total of 2514 articles were identified in the initial search.

“A total of 2514 articles were identified in the initial search.”

• Also stated in: § Overview · plain

  “The initial search from MEDLINE, IEEE Xplore, and ScienceDirect provided cumulative results of 2514 publications.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

After abstract review, 134 (27.7%) publications were identified for further text review, excluding 345 (71.4%) publications, including 4 (0.8%) duplicate articles.

“After abstract review, 134 (27.7%) publications were identified for further text review, further excluding 345 (71.4%) publications, including 4 (0.8%) duplicate articles.”

May combine two claims.

During the initial search, a majority of the literature seemed to report on a specific area, such as radiology, in a specific setting, and a number of studies were reporting on implementation findings and not necessarily adoption.

“During this search, it was found that a majority of the literature seemed to report on a specific area, such as radiology, in a specific setting (hospital or community), and a number of studies were reporting on implementation findings and not necessarily adoption.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The themes were found to provide perspective on both facilitators and barriers.

“Interestingly, the themes were found to provide perspective on both facilitators and barriers.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Other studies noted that the way the data were being used to train algorithms was not clear.

“Other studies noted that the way the data were being used to train algorithms was not clear.”

May combine two claims.

The majority of the articles discussed AI in health care in general, with a majority reporting from the field of radiology or oncology.

“In terms of setting, the majority of the articles discussed AI in health care in general with a majority of the articles reporting from the field of radiology or oncology.”

Probably not a substantial claim: may not pass S3 (load-bearing).

After screening, 483 (19.21%) publications were included for abstract review based on the title of the study.

“After screening the results, 483 (19.21%) publications were included for abstract review based on the title of the study.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The reporting of results for each theme provided perspective on the theme being both a barrier and facilitator, with the exception of governance, which was entirely noted as a key facilitator.

“As such, the reporting of the results for each theme provided perspective on the theme being both a barrier and facilitator, with the exception of governance, which was entirely noted as a key facilitator.”

May combine two claims.

The lack of traceability and logical understanding of how the algorithm arrived at a recommendation contradicted a key foundation of evidence-based medicine, which relies on high standards of explainability.

“The lack of traceability and logical understanding of how the algorithm arrived at a recommendation contradicted a key foundation of evidence-based medicine, which relies on high standards of explainability.”

Probably not a substantial claim: may not pass S3 (load-bearing).

AI not accounting for patients' health determinants was noted as a "grand challenge."

“AI not accounting for patients’ health determinants was noted as a “grand challenge” [19].”

• cites [19] sourceNot checked — no full text of the source

May combine two claims.

Lack of sufficient buy-in from end users and lack of endorsement from organizational leadership emerged as barriers, including not engaging stakeholders early in the process.

“A lack of sufficient buy-in from end users and a lack of endorsement from organizational leadership emerged as barriers, including not engaging stakeholders early in the process.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Adoption of IBM Watson in China faced challenges due to patients having to pay high fees for the service.

“Sun and Medaglia [50] additionally noted challenges associated with the adoption of IBM Watson in China due to patients having to pay high fees for the service.”

• cites [50] sourceNot checked — no full text of the source

May combine two claims.

Algorithm bias is a critical factor in not only gaining trust but also having meaningful outputs that can be applied to diverse patients.

“In terms of equity and fairness, our findings have demonstrated that algorithm bias is a critical factor in not only gaining trust but also having meaningful outputs that can be applied to diverse patients.”

May combine two claims.

An artificial vision algorithm was unable to identify dark-skinned individuals, as >80% of the individuals in a reference data set were light-skinned individuals.

“Buolamwini and Gebru [85] reported that an artificial vision algorithm was unable to identify dark-skinned individuals, as >80% of the individuals in a reference data set were light-skinned individuals.”

• cites [85] BuolamwiniJGebruTGender shades: intersectional accuracy disparities in commercial gender classificationProceedings of the 2018 International Conference on Machine Learning and Machine Intelligence2018MLMI '18September 28-30, 2018Ha Noi, VietNam115Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The framework developed by Landers and Behrend comprehensively outlines questions that would be asked at each stage of AI development.

“Frameworks such as the one developed by Landers and Behrend [91] comprehensively outline questions that would be asked at each stage of AI development.”

• cites [91] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Numerous solutions to address bias emphasize applying bias mitigation techniques at each stage of model development.

“There are numerous solutions to address bias that emphasize the risk of bias mitigation techniques to be applied at each stage of model development.”

May combine two claims.

Common themes identified around barriers to adoption included explainability, trust issues centered on privacy, challenges around data ownership, lack of regulatory standards, issue of bias, and lack of accountability.

“Common themes identified around barriers to adoption included explainability, trust issues centered on privacy, challenges around data ownership, lack of regulatory standards, issue of bias, and lack of accountability.”

May combine two claims.

The issue of trust was found to be centered on bias, ethics, and explainability, which led to a lack of accountability and an inability to evaluate.

“Overall, the issue of trust was found to be centered on bias, ethics, and explainability, which led to a lack of accountability and an inability to evaluate.”

Classified as probably not a substantial claim.

A thematic analysis was conducted on the selected articles to map factors associated with barriers to and facilitators of AI adoption in health care.

“A thematic analysis was conducted on the selected articles to map factors associated with the barriers to and facilitators of AI adoption in health care.”

Probably not a substantial claim: may not pass S4 (based).

The effectiveness of AI in health care relies on successful implementation and adoption by clinicians, patients, and other health care stakeholders.

“However, the translation of AI into practical, everyday use has been limited, as its effectiveness relies on successful implementation and adoption by clinicians, patients, and other health care stakeholders.”

Classified as probably not a substantial claim.

MEDLINE, IEEE Xplore, and ScienceDirect databases were searched to identify publications in English reporting barriers to or facilitators of AI adoption in health care.

“MEDLINE, IEEE Xplore, and ScienceDirect databases were searched to identify publications in English that reported on the barriers to or facilitators of AI adoption in health care.”

Classified as probably not a substantial claim.

A governance structure can be a key facilitator in ensuring all the elements identified as barriers are addressed appropriately.

“A governance structure can be a key facilitator, among others, in ensuring all the elements identified as barriers are addressed appropriately.”

• Also stated in: § Conclusions · plain

  “A governance structure can be a key facilitator in ensuring all the elements identified as barriers are addressed appropriately.”

Classified as probably not a substantial claim. May combine two claims.

A conventional content analysis approach was used to review the articles, chart the data, and identify themes.

“A conventional content analysis approach was used to review the articles, chart the data, and identify themes [30].”

• cites [30] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Any discrepancies in screening were resolved through discussion.

“Any discrepancies were resolved through discussion.”

May combine two claims.

Both the unified theory of acceptance and use of technology and the nonadoption, abandonment, scale-up, spread, and sustainability frameworks emphasize the importance of studying adoption to support the successful uptake of technologies beyond implementation.

“Both emphasize the importance of studying adoption to support the successful uptake of technologies beyond implementation.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Two reviewers independently screened the articles from the initial search by reviewing their titles and abstracts.

“A total of 2 reviewers independently screened the articles from the initial search by reviewing their titles and abstracts.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Duplicate studies were identified and removed to ensure there was no overlap.

“Duplicate studies were identified and removed to ensure there was no overlap.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The unified theory of acceptance and use of technology framework suggests several factors that help understand users' intention to adopt and use a technology.

“For example, the unified theory of acceptance and use of technology framework, which integrates all the available theories about technology adoption, suggests several factors that help understand users’ intention to adopt and use a technology.”

• cites [23] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Bias can be addressed by engaging clinicians in the design and development of the algorithm to ensure that appropriate measures are taken to address bias before the AI algorithm is deployed.

“This can be addressed by engaging clinicians in the design and development of the algorithm to ensure that appropriate measures are taken to address bias before the AI algorithm is deployed.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The opacity of the output is not aligned with the "medical ethos."

“Similarly, Gillner [44] also noted that this opacity of the output is not aligned with the “medical ethos.””

• cites [44] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Four primary ethical challenges need to be addressed to realize the full potential of AI in health care: informed consent to use, safety and transparency, algorithmic fairness and biases, and data privacy.

“Gerke et al [92] have discussed four primary ethical challenges that need to be addressed to realize the full potential of AI in health care: (1) informed consent to use, (2) safety and transparency, (3) algorithmic fairness and biases, and (4) data privacy.”

• cites [92] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Goirand et al note that operationalizing ethical frameworks for AI is challenging, as there is a need for contextualization due to different ethical issues present in different environments.

“Despite these practical approaches, Goirand et al [98] note that operationalizing ethical frameworks for AI is challenging, as there is a need for contextualization due to different ethical issues present in different environments.”

• cites [98] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable).

The existing technology implementation and acceptance models may not be all encompassing of adoption factors.

“The existing technology implementation and acceptance models may not be all encompassing of adoption factors”

Probably not a substantial claim: may not pass S3 (load-bearing).

From a clinician's perspective, there are concerns around what the clinician's responsibility is in informing patients about the use of AI in their care.

“From a clinician’s perspective, there are concerns around what the clinician’s responsibility is in informing patients about the use of AI in their care [92].”

• cites [92] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable).

Adding additional frameworks around trust, bias, explainability, and ethics will be necessary to foresee the success of an AI innovation.

“adding additional frameworks around trust, bias, explainability, and ethics will be necessary to foresee the success of an AI innovation.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

These findings around trust resonate with results from this research, reinforcing the barriers to adoption identified in both studies.

“These findings around trust resonate with results from this research, reinforcing the barriers to adoption identified in both studies.”

• cites [27] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

At the time this search was conducted, a few literature reviews on the determinants of and barriers to AI adoption were conducted, such as the review by Radhakrishnan and Chattopadhyay.

“At the time this search was conducted, a few literature reviews on the determinants of and barriers to AI adoption were conducted, such as the review by Radhakrishnan and Chattopadhyay [26].”

• cites [26] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

A governance model may address concerns around risk, safety, and adoption barriers identified in this paper by facilitating the overall development process of AI and ensuring various checks and balances are in place.

“A governance model may address concerns around risk, safety, and adoption barriers identified in this paper by facilitating the overall development process of AI and ensuring various checks and balances are in place.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Given the limited application of AI in health care at the time of this research, only a few number of papers that reported on implementation barriers and facilitators were reviewed to identify AI adoption barriers and facilitators.

“Given the limited application of AI in health care at the time of this research, only a few number of papers that reported on implementation barriers and facilitators were reviewed to identify AI adoption barriers and facilitators.”