1. Introduction

 

Artificial Intelligence is increasingly embedded in healthcare systems, supporting clinical decision-making through advanced data analytics, machine learning, and predictive modelling. These technologies enable clinicians to process complex datasets, improve diagnostic accuracy, and enhance treatment outcomes, with demonstrated success in radiology, oncology, and chronic disease management [1,2]. However, the integration of AI into clinical environments raises significant ethical and governance concerns. Healthcare decisions are inherently value-driven, involving patient autonomy, safety, equity, and trust. The introduction of AI systems - often opaque, data-dependent, and dynamically updating - challenges traditional ethical frameworks and professional practices.

Global institutions, including the World Health Organization (WHO) [3], emphasise that AI in healthcare must be developed and deployed with ethics and human rights at its core. The WHO identifies key risks such as bias, lack of transparency, and accountability gaps, highlighting the need for structured governance mechanisms. Similarly, the European Commission's AI Act [4] classifies many medical AI systems as high-risk, mandating rigorous oversight. Early and foundational work by Vayena et al. [5] identified three core ethical challenges for machine learning in medicine: data sourcing and privacy, algorithmic fairness and bias, and transparency and accountability in clinical deployment. This recognition was extended by Char et al. [6], who warned that machine learning systems could perpetuate racial bias, lack transparency, and strain the fiduciary relationship between physicians and patients if implemented without appropriate safeguards.

Despite regulatory progress at the supranational level, empirical evidence reveals a persistent implementation gap. A 2026 WHO European Region survey of all 27 EU Member States - with a 94% response rate across the wider Region - found that only 11% have adopted health-specific AI strategies, only 7% have issued health-specific ethical guidelines, and just 11% have established dedicated legal liability frameworks for AI in healthcare [7]. Furthermore, only 24% of countries have agencies responsible for post-market monitoring of AI systems, and merely 26% offer in-service AI training for health professionals.

Conversely, advanced exemplars show what is possible. The World Economic Forum [8] white paper on Abu Dhabi's intelligent health system describes a large-scale initiative integrating over 100,000 clinical, genomic, financial, and environmental data streams into a unified Population Health Intelligence (PHI) platform. The system has achieved measurable improvements: reducing heart attack response times to 57 minutes, supporting earlier breast cancer risk assessment and preventive screening strategies, and detecting insurance fraud in real time. As a model for other health systems, Abu Dhabi's approach is exemplary. At the same time, like any pioneering and rapidly evolving intelligent health ecosystem, the case also highlights opportunities for continued advancement in areas such as clinician override governance, transparent patient communication frameworks, and structured bias auditing processes.

This paper addresses the gap between ethical principles and operational implementation by: (1) systematically reviewing ethical challenges in AI-driven clinical decision-making across five domains; (2) presenting comparative empirical evidence from 27 EU Member States; (3) analysing a high-performing exemplar (Abu Dhabi) to identify remaining gaps; and (4) proposing a three-level governance framework with specific, actionable mechanisms validated against both cross-national data and a real-world case.

2. Methodology

2.1 Search Strategy

This systematic review was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement. A systematic literature review was conducted across PubMed, Scopus, and Web of Science. The search was conducted between January and April 2026. Keywords combined MeSH terms and free text: ("artificial intelligence" or "machine learning" or "AI") and ("clinical decision-making" or "healthcare") and  ("ethics" or "algorithmic bias" or "accountability" or "informed consent" or "governance"). Hand-searching of reference lists and key journals (Journal of Medical Ethics, The Lancet Digital Health) was performed.

 

2.2 Inclusion and Exclusion Criteria

Inclusion: Peer-reviewed articles or official organisational reports (WHO [3], OECD, UNESCO [27], World Economic Forum [8]); published 2018–2026; English language; explicit focus on AI in clinical decision-making with substantial ethical or governance discussion.

Exclusion: Non-healthcare AI studies; purely technical papers without ethical analysis; editorials or opinion pieces without original synthesis.

Records were screened by title and abstract; full texts of potentially relevant records were then assessed against the inclusion criteria.

Initial database searches identified approximately 180 records across PubMed, Scopus, and Web of Science. After removal of duplicates, 135 records remained for title and abstract screening. Full-text assessment was conducted for potentially relevant studies, and 27 sources met the final inclusion criteria. Studies were excluded if they focused solely on technical AI development without ethical analysis, non-healthcare applications, or lacked relevance to clinical decision-making governance.

Notable included sources: WHO European Region 2026 survey report providing empirical data from all 27 EU Member States and the World Economic Forum [8] white paper presenting a detailed case study of a high-performing intelligent health system.

2.3 Analysis

Thematic analysis was conducted to identify recurrent ethical themes across the literature. The author coded all included sources line-by-line. Initial codes (e.g., 'bias in data', 'lack of transparency', 'clinician over-reliance') were grouped through iterative review into broader candidate themes. These were then refined and defined, yielding five final ethical domains: algorithmic bias, physician autonomy, informed consent, accountability, and the patient–provider relationship. These five domains were selected because they emerged most consistently across the peer-reviewed literature and policy reports. Algorithmic bias appeared in discussions of fairness and health equity; physician autonomy in debates about automation and clinical judgment; informed consent in relation to transparency and explainability; accountability in legal and regulatory frameworks; and the patient–provider relationship in concerns about trust, empathy, and shared decision-making. Together, these five domains capture the core ethical challenges of AI in clinical decision-making identified in the literature. A second coding pass was conducted after a two-week interval to check internal consistency. Governance gaps were identified through cross-study comparison. The WHO Europe survey data were used to quantify the prevalence of governance mechanisms across 27 countries. The Abu Dhabi case was used to test the relevance of each domain and to identify remaining gaps in an advanced system.

2.4 Data Synthesis

Quantitative findings from the WHO Europe survey [7] are presented in tables and figures. Qualitative findings from peer-reviewed literature and the WEF case study are integrated thematically.

3. Results

 

The analysis of peer-reviewed literature and policy reports revealed five interconnected ethical domains: algorithmic bias, physician autonomy, informed consent, accountability, and the patient–provider relationship. Table 1 presents key governance indicators from the WHO Europe survey [7] that inform and support these five ethical domains, demonstrating how specific empirical gaps correspond to normative ethical deficits. This mapping provides the analytical bridge between descriptive findings and the prescriptive framework proposed in Section 6.

 

Table 1. Key Governance Indicators Across 27 EU Member States [7]

Governance Domain	Indicator	% (n/N)
National AI Strategies	Health-specific AI strategy adopted	11% (3/27)
	Health-specific AI strategy under development	15% (4/27)
	Cross-sectoral AI strategy adopted	85% (23/27)
Ethical Guidelines	Health-specific ethical guidelines issued	7% (2/27)
	Cross-sectoral ethical guidelines issued	19% (5/27)
	No ethical guidelines issued	44% (12/27)
Legal Liability	Dedicated AI liability framework established	11% (3/27)
	Planning alignment with EU legislation	56% (15/27)
Regulatory Oversight	Agencies for AI approval (pre-market)	56% (15/27)
	Agencies for post-market monitoring	24% (12/50 Region)
Workforce Capacity	In-service AI training offered	26% (7/27)
	Pre-service AI training offered	22% (6/27)
	New AI/data science roles created	48% (13/27)
Stakeholder Engagement	Public consulted on AI governance	18% (4/22)
	Patient associations consulted	36% (8/22)
	Consultation insights made public	32% (7/22)
Data Governance	National health data hub established	63% (17/27)
	Health data strategy adopted	67% (18/27)
Barriers	Financial affordability as major barrier	41% (11/27)
	Legal uncertainty as major barrier	33% (9/27)
	Data quality/standards as major barrier	33% (9/27)
Policy Enablers	Transparency guidance (major positive impact)	63% (17/27)
	Liability rules (major positive impact)	56% (15/27)
	Legal guidance on data use (major positive impact)	56% (15/27)

Source: Author's synthesis of data from WHO Regional Office for Europe [7], Artificial intelligence is reshaping health systems: state of readiness across the European Union, sections 3.1–3.6. Notes: N = total number of EU Member States that answered the relevant survey question. For most indicators, N = 27. For rows with N = 22, the question was asked only of countries that had conducted stakeholder engagement. For the row marked 'Region', N = 50 refers to the wider WHO European Region.

 

Table 1b. Mapping of WHO Europe Governance Indicators to Five Ethical Domains

 

Ethical Domain	Corresponding WHO Europe Indicator (from Table 1)	Justification
Algorithmic Bias	Data quality/standards as major barrier (33% of EU Member States); Genomic data included in only 18% of national health data hubs	Poor data representativeness and quality directly produce algorithmic bias [9,10]
Physician Autonomy	In-service AI training offered (26%); New AI/data science roles created (48%)	Without adequate training, clinicians cannot critically evaluate AI recommendations, increasing risk of automation bias [12]
Informed Consent	Transparency guidance rated as major positive impact (63%) but health-specific ethical guidelines issued (only 7%); Public consulted on AI governance (18%)	Informed consent requires transparency and patient understanding, which depend on clear guidelines and public engagement [14]
Accountability	Dedicated AI liability framework established (11%); Agencies for post-market monitoring (24% in Region)	Clear liability rules and post-market surveillance are core accountability mechanisms [17]
Patient–Provider Relationship	Patient associations consulted (36%); Public consulted (18%); Trust rated as important (93% of Member States)	Trust and patient engagement are foundational to the therapeutic relationship; lack of consultation undermines both [18]

Source: Author’s mapping of WHO Regional Office for Europe [7] governance indicators to ethical domains identified in systematic review.

Table 1b explicitly maps each governance indicator from Table 1 to the ethical domain it supports, providing the logical bridge between empirical observations and normative ethical categories.

 

 

Figure 1: AI Governance Mechanisms Across EU Member States [7]   Health-specific AI strategy       ░░  11% Health-specific ethical guidelines ░   7% Dedicated liability framework      ░░  11% Post-market monitoring agency      ░░░  24% In-service AI training             ░░░  26% Public consultation on AI          ░░  18% National health data hub           ████████████████████  63% Cross-sectoral AI strategy         ████████████████████████████  85%   0%      20%      40%      60%      80%     100%

Figure 1. AI Governance Mechanisms Across EU Member States [7]. Source: Author's synthesis; WHO Regional Office for Europe [7], sections 3.1–3.6.

 

3.1 Algorithmic Bias and Health Inequality

AI systems trained on historical datasets may reflect or amplify existing healthcare disparities. Obermeyer et al. [9] demonstrated that a widely used US healthcare algorithm systematically underestimated the medical needs of Black patients compared to White patients with equivalent illness burden. Similarly, Seyyed-Kalantari et al. [10] found that commercial chest X-ray algorithms performed significantly worse on underrepresented racial groups. Chen et al. [11] further argue that algorithmic discrimination often stems from inadequate sample sizes or unmeasured predictive variables, and that targeted data collection - rather than model constraints alone - can reduce unfairness without sacrificing accuracy.

Cross-country evidence [7]: Data quality and standards were ranked as the third most important barrier to AI adoption, with 33% of EU Member States rating it as of major importance. However, only 63% of countries have established national health data hubs, and genomic data are included in just 18% of these hubs - highlighting persistent data representativeness gaps.

Exemplar case [8]: The Abu Dhabi Emirati Genome Programme has sequenced over 850,000 samples, addressing the historic underrepresentation of Arab populations in global genomic datasets. Given Abu Dhabi's leadership in this space, it is well-positioned to pioneer the next frontier: mandatory bias auditing. Currently, the WEF report does not specify mandatory bias auditing schedules stratified by demographic variables. It is important to note that absence from the report does not confirm absence in practice; however, the lack of documented governance mechanisms represents an opportunity for further advancement that this framework addresses.

3.2 Physician Autonomy and Automation Bias

AI systems influence clinical decision-making, potentially leading to automation bias - the tendency to over-rely on algorithmic outputs without critical scrutiny. Grote & Berens [12] argue that institutional pressures can push clinicians to follow AI recommendations even when professional judgment suggests otherwise. Blease et al. [13] further note that this gradual shift may erode clinical reasoning skills over time.

Cross-country evidence [7]: Reducing pressure on the healthcare workforce was ranked as the second most important driver for AI adoption, with 96% of EU Member States rating it as major or moderate relevance. However, only 26% offer in-service AI training for health professionals, and only 48% have created new AI/data science roles. This training gap directly increases the risk of automation bias.

Exemplar case [8]: The Abu Dhabi Unified Medical Operations Centre (UMOC) uses AI for real-time emergency response, route optimisation, and capacity management, reducing heart attack response times to 57 minutes. However, the WEF report does not specify protocols for when and how clinicians can override AI recommendations in high-pressure settings - a critical gap for preserving physician autonomy.

3.3 Informed Consent and Transparency

Traditional informed consent assumes patients understand the basis of medical decisions. However, many AI systems operate as "black boxes" with limited explainability. Floridi et al. [14] and Holzinger et al. [15] argue that meaningful informed consent requires at least a basic explanation of how AI contributes to diagnosis or treatment. However, Ghassemi et al. [16] caution that current explainability methods are unlikely to achieve these goals for patient-level decision support, advocating instead for rigorous internal and external validation of AI models as a more direct means of building trust. 

Cross-country evidence [7]: Transparency, verifiability and explainability of AI solutions was rated as the highest-impact policy enabler, with 63% of EU Member States ranking it as having major positive impact. Yet only 7% have issued health-specific ethical guidelines addressing transparency, and only 18% have consulted the broader public on AI governance. This disconnect is illustrated in Figure 2.

Exemplar case [8]: Abu Dhabi's Sahatna application provides personalised screening prompts based on AI analysis. However, the WEF report does not detail patient consent processes for AI-assisted decisions or provide accessible mechanisms for patients to question or opt out of AI-driven suggestions.

 

 

Figure 2: Disconnect Between Policy Enabler Importance and Actual Adoption                                   Perceived as        Actually                                 major positive      adopted                                 impact Transparency guidance:          63%           vs    7%  health-specific guidelines Liability rules:                56%           vs   11%  dedicated frameworks Data use guidance:              56%           vs   44%  data sharing rules w/ private sector

Figure 2. Disconnect Between Policy Enabler Importance and Actual Adoption [7]. Source: Author's synthesis; WHO Regional Office for Europe [7], sections 3.3 and 3.6.

 

3.4 Accountability and Legal Responsibility

AI introduces a diffusion of responsibility among clinicians, developers, healthcare institutions, and regulators. Gerke et al. [17] observe that conventional medical malpractice frameworks assume human intent or negligence, which does not translate easily to autonomous learning systems. Morley et al. [22] describe an 'accountability gap': when an AI-generated recommendation leads to patient harm, no single actor may be clearly liable.

Cross-country evidence [7]: Accountability and liability rules were rated as having major positive impact by 56% of EU Member States. However, only 11% have established dedicated legal liability frameworks for AI in healthcare. Table 2 summarises the liability landscape.

Exemplar case [8]: Abu Dhabi's Shafafiya platform uses AI to monitor insurance claims for fraud in real time. However, the WEF report does not address liability when AI-driven claim denials or clinical recommendations lead to patient harm - reinforcing the accountability gap.

 

Table 2. Liability and Oversight Mechanisms Across EU Member States [7]

Mechanism	Percentage	Status
Dedicated AI liability framework established	11%	Adopted
New AI-specific liability framework	4%	Adopted (Belgium)
Guidance on applying existing liability laws	7%	Adopted
Developing new liability standards	11%	In development
Planning alignment with EU legislation	56%	Planned
Agencies for pre-market AI approval	56%	In place
Agencies for post-market monitoring	24%	In place (Region)

Source: Author's synthesis of data from WHO Regional Office for Europe [7], sections 3.3 and 3.4.

 

3.5 Patient–Provider Relationship

AI can both enhance and undermine the patient–provider relationship. While AI has the potential to improve diagnostic precision and efficiency, concerns remain that excessive reliance on technology may reduce empathy, human interaction, and the therapeutic dimension of care [19]. Concerns have been raised that opaque and poorly explainable artificial intelligence systems may reduce trust in clinical decision-making and create challenges for transparency and accountability in patient care [16]. Conversely, when implemented appropriately, AI-enabled decision aids can support shared decision-making, improve patient engagement, and facilitate more personalised communication between patients and healthcare providers [20].

Cross-country evidence [7]: Trust was rated as having at least some importance by 93% of EU Member States. However, patient associations were consulted in only 36% of countries, and the broader public in only 18%. This lack of patient and public engagement risks developing AI tools that do not align with patient values.

Exemplar case [8]: An example of population-level implementation is the Abu Dhabi Healthy Living Strategy, which used Population Health Intelligence (PHI) insights to support more than 200 community health activations in high-risk districts [8]. While AI-driven public health initiatives can improve targeting and efficiency, researchers have cautioned that insufficiently human-centred implementation may negatively affect perceptions of personalisation, trust, and patient engagement [21,16].

3.6 Summary of Key Findings

3.6.1 From Empirical Indicators to Ethical Domains: Justification of the Mapping

The five ethical domains identified in this review (algorithmic bias, physician autonomy, informed consent, accountability, and patient–provider relationship) were derived from thematic analysis of peer-reviewed literature. However, the relationship between the WHO Europe [7] governance indicators presented in Table 1 and these ethical domains requires explicit justification to avoid interpretive leaps.

Table 1b above maps each governance indicator to its corresponding ethical domain and explains the logical connection. For example:

·       Low rates of in-service AI training (26%) directly undermine physician autonomy by leaving clinicians unprepared to critically evaluate AI recommendations, increasing the risk of automation bias [12].

·       The absence of public consultation on AI governance (only 18% of countries) violates the transparency requirement for informed consent, as patients cannot meaningfully consent to AI-assisted decisions without understanding how these systems operate [14].

·       The lack of dedicated AI liability frameworks (only 11%) creates an accountability gap, as no clear responsibility is assigned when AI systems cause harm [17].

·       Poor data quality and representativeness (33% cite as major barrier; genomic data in only 18% of hubs) directly fuels algorithmic bias, as AI systems trained on non-representative data systematically underperform for underrepresented groups [9,10].

·       Low rates of patient and public consultation (36% and 18% respectively) undermine the patient–provider relationship by excluding patient voices from AI governance decisions that affect trust and communication [18].

This mapping clarifies how empirical governance gaps translate into normative ethical deficits, providing the analytical bridge between the descriptive findings of the WHO Europe survey and the prescriptive framework proposed in Section 6.

Table 3 provides a consolidated summary of the governance gap across all five ethical domains, comparing the evidence from peer-reviewed literature, WHO Europe cross-country data, and the Abu Dhabi exemplar.

 

Table 3. Governance Gap Across Five Ethical Domains — Triangulated Evidence

Ethical Domain	Peer-Reviewed Evidence	WHO Europe [7] Evidence	Abu Dhabi [8] Evidence	Gap Identified
Algorithmic Bias	9,10	33% cite data quality as major barrier; genomic data in only 18% of hubs	Genome programme addresses representation; mandatory bias audits not documented in the WEF report	No systematic, stratified bias auditing
Physician Autonomy	12,13	96% cite workforce pressure as driver; only 26% offer training	UMOC reduces response times; override protocols not specified in the WEF report	No human-in-the-loop documentation
Informed Consent	14,15	63% rate transparency as top enabler; 7% have guidelines	Sahatna app provides insights; consent scripts not detailed in the WEF report	No standardised patient communication
Accountability	17,22	56% rate liability as key enabler; 11% have frameworks	Shafafiya detects fraud; liability pathways not specified in the WEF	No clear legal or institutional responsibility
Patient–Provider Relationship	16; 13,20	93% rate trust as important; 18% consult public	Healthy Living Strategy; risk of depersonalisation noted (general caution from literature, not criticism of Abu Dhabi)	No mandated patient engagement

Source: Author's synthesis triangulating: (1) peer-reviewed literature; (2) WHO Regional Office for Europe [7], sections 3.1–3.6; (3) World Economic Forum [8], A New Era for Digital Health: Abu Dhabi's Leap to Health Intelligence.

 

4. Discussion: From Ethics to Implementation

 

While ethical principles are well established, their practical implementation remains limited. The WHO [3] identifies six core principles for AI in healthcare: protection of human autonomy, promotion of well-being and safety, transparency and explainability, accountability, inclusiveness and equity, and sustainability. However, healthcare systems often adopt AI technologies faster than governance frameworks evolve. Recent scholarship has further emphasised the need for internationally coordinated governance structures to ensure responsible implementation of AI and data-driven healthcare systems across jurisdictions [23]. Lyell et al. [24] found that many hospitals deployed AI systems before establishing internal ethical governance structures, creating what we term a governance gap: ethics exist conceptually, but implementation is fragmented, reactive, and inconsistent.

Sendak et al. [25] similarly observed that despite enormous enthusiasm, machine learning models are rarely translated into clinical care with minimal evidence of clinical or economic impact. Drawing on 21 case studies of products integrated into routine care, they identified a four-phase translational path (design, evaluate, scale, monitor) and highlighted challenges including lack of interoperability, stealth science, and insufficient post-deployment surveillance - gaps that mirror the ethical governance gaps identified in this review. Van de Sande et al. [26] conducted a systematic review of 494 AI studies in the intensive care unit and found that 89.3% of models never progressed beyond prototyping, 80.9% of retrospective studies had high risk of bias, and critically, no studies reported on AI models integrated into routine clinical practice - a finding that starkly illustrates the gap between AI development and real-world implementation.

The WHO Europe [7] survey provides the first comprehensive empirical quantification of this governance gap across 27 EU Member States. Key findings: only 11% have health-specific AI strategies; only 7% have health-specific ethical guidelines; only 11% have dedicated AI liability frameworks; only 26% offer in-service AI training; and only 18% consult the broader public on AI governance.

The Abu Dhabi case [8] serves as an exemplary model of what is possible when political will, investment, and data infrastructure converge. The emirate has built interoperable data architecture, integrated over 100,000 clinical, genomic, financial, and environmental data streams, and achieved measurable clinical improvements — reducing heart attack response times to 57 minutes (well below the global 90-minute benchmark) and supporting earlier breast cancer risk assessment and preventive screening strategies — demonstrating that intelligent health systems are achievable today. Precisely because Abu Dhabi is at the forefront of this transformation, its experience offers valuable insights for further advancement. The WEF white paper, while comprehensive in describing technical infrastructure and outcomes, does not document governance mechanisms such as clinician override protocols, transparent patient communication frameworks, liability allocation pathways, and structured bias auditing processes stratified by demographic variables. It is important to clarify that absence of documentation in the public report does not confirm absence of these mechanisms in Abu Dhabi's actual practice; rather, it indicates that these governance details are outside the scope of that report. Abu Dhabi's system remains an exemplary model for data integration and clinical outcomes.

Having established the empirical evidence for the governance gap across five ethical domains, the paper now transitions from descriptive findings to a prescriptive framework. The following section proposes specific operational mechanisms to address the identified gaps. This separation ensures that empirical observations (what is) are clearly distinguished from normative recommendations (what should be).

Table 4 operationalises the concept using algorithmic bias as a case example, showing that while technical capability and ethical acceptance exist in most systems, every operational requirement for systematic, stratified bias auditing is absent.

 

Table 4. Why Is This a 'Governance Gap'? - Algorithmic Bias as a Case Example

A governance gap exists when technical capability and ethical acceptance are present, but mandatory operational mechanisms are absent.

Element	Status in Most Systems (incl. Abu Dhabi)	What 'Systematic, Stratified Auditing' Would Require	Gap Status
Technical capability	✅ Yes (integrated data, stratified analysis possible)	Use existing capability	Operationally established
Ethical principle accepted	✅ Yes (WHO, EU AI Act, Abu Dhabi endorse fairness)	Translate principle into enforceable requirement	Operationally established
Mandatory pre-deployment audit	❌ Not formally specified	Every AI system must pass bias test before clinical use	Operational governance gap
Annual re-audit	❌ Not formally specified	Bias retested annually as population/data evolve	Operational governance gap
Stratification by age/sex/race/SES	❌ Not formally specified	Performance reported separately for each subgroup	Operational governance gap
Public reporting	❌ Not formally specified	Audit results accessible to regulators, ethics committees, and public	Operational governance gap
Enforcement (consequences for failure)	❌ Not formally specified	System cannot be deployed or must be retrained if bias exceeds threshold	Operational governance gap

Source: Author's original framework. Status column supported by: Obermeyer et al. [9]; Seyyed-Kalantari et al. [10]; WHO Europe [7], sections 3.1–3.4; WEF Abu Dhabi 2026. Notes:The 'Status in Most Systems (incl. Abu Dhabi)' column is based on documentation available in the [8] white paper, which focuses on technical infrastructure and clinical outcomes. The absence of documented requirements in the white paper does not imply that such mechanisms are absent in Abu Dhabi's actual practice; rather, it indicates that these governance details are not within the scope of that report. Abu Dhabi's system remains an exemplary model for data integration and clinical outcomes.

 

As Table 4 demonstrates, the gap is not in what is possible or what is agreed upon, but in what is mandated, operationalised, and enforced. Technical capability exists. Ethical principles are endorsed. Yet the operational mechanisms - mandatory pre-deployment audits, annual re-audits, stratification by demographic variables, public reporting, and enforcement consequences - are absent across most systems, including advanced exemplars.

This pattern - capability without mandate - recurs across all five ethical domains identified in this review:

 

Ethical Domain	Technical Capability Exists?	Ethical Principle Accepted?	Mandatory Operational Mechanism Exists?
Algorithmic Bias	Yes	Yes	No systematic, stratified audit
Physician Autonomy	Yes (override logs possible)	Yes (human oversight endorsed)	No mandated override documentation
Informed Consent	Yes (explainable AI exists)	Yes (transparency endorsed)	No standardised patient scripts
Accountability	Yes (audit trails possible)	Yes (liability principles exist)	No clear legal framework
Patient–Provider Relationship	Yes (engagement platforms exist)	Yes (trust endorsed)	No mandated patient engagement

 

The implication is clear: moving from ethical principles to accountable AI requires not just more technology or more guidelines, but binding operational requirements embedded at every level of the health system.

 

5.  Limitations

This study has several limitations that should be acknowledged.

First, the analysis of the Abu Dhabi intelligent health system relies on a single publicly available white paper [8], which focuses on technical infrastructure and clinical outcomes rather than governance protocols. As noted in Section 4, absence of documented mechanisms in this report does not confirm their absence in practice.

Second, the systematic review was limited to English-language peer-reviewed articles and policy reports, potentially excluding relevant studies published in other languages. This may introduce language bias.

Third, the governance indicators from the WHO Europe survey [7] are based on self-reported data from Member States, which may be subject to social desirability bias or variations in interpretation across countries.

Fourth, the rapidly evolving regulatory landscape for AI in healthcare — including the ongoing implementation of the EU AI Act [4] and the European Health Data Space regulation — means that some findings may become outdated as new regulations take effect. The governance framework proposed in this paper is designed to be adaptive, but specific policy recommendations may require revision as the regulatory environment matures.

Fifth, this review was conducted by a single author. While a second coding pass was performed after a two-week interval to enhance internal consistency, independent dual coding was not possible. Future research should include multiple reviewers to strengthen reliability.

Finally, the proposed three-level governance framework has not been empirically tested in real-world settings. The framework represents a normative recommendation based on identified gaps; empirical validation remains an important direction for future research.

 

 

6. Proposed Three-Level Ethical Governance Framework for Clinical AI

 

Building on the empirical findings presented in Sections 3 and 4, this section now proposes a normative governance framework. The framework is prescriptive in nature, outlining what should be implemented to address the identified gaps. It is informed by global ethical standards [3,4] and contextualised through the Abu Dhabi exemplar [8], but the mechanisms themselves are recommendations, not empirical observations.

The following framework is normative, not descriptive. The empirical evidence for governance gaps was established in Sections 3 and 4. The mechanisms proposed below are recommendations based on that evidence, not findings from the data. Where empirical evidence supports a specific recommendation (e.g., low rates of AI training justify mandatory training programmes), this is noted. However, the framework itself represents an original contribution, not a direct extrapolation from any single data source.

 

Table 5. Three-Level Ethical Governance Framework for AI in Clinical Decision-Making

Level	Key Mechanisms	Responsible Actors	WHO Europe Gap Addressed	Abu Dhabi Gap Addressed
System-Level (Policy & Regulation)	Mandatory alignment with WHO/EU standards; pre-market conformity assessments for bias; enforceable national guidelines	National governments; regulators; EU institutions	Only 11% have health-specific strategies; 7% have ethical guidelines	Opportunity for formalised bias auditing governance
Organisational-Level (Hospitals & Institutions)	AI ethics committee (multidisciplinary); mandatory bias auditing (annual, stratified); data governance policies; clinician AI literacy training	Hospital boards; AI ethics committees; data governance officers	Only 26% offer in-service training; 48% have new roles; 24% have post-market monitoring	Opportunity for formalised ethics oversight and structured audit schedules
Clinical-Level (Practice & Decision-Making)	Human-in-the-loop with override documentation; transparent patient communication protocols; AI as decision aid, not substitute	Clinicians; patients; clinical teams	18% consult public; 36% consult patient associations; no consent standards	Opportunity for enhanced clinician oversight and patient communication standardisation

Source: Author's original governance framework developed from the synthesis of evidence in this review. Informed by: (1) WHO [3] and EU AI Act [4]; (2) WHO Europe [7] governance gaps (Table 1); (3) WEF Abu Dhabi 2026 gaps (Table 3); and (4) operationalised gap analysis (Table 4).

 

6.1 System-Level Governance (Policy & Regulation)

Mechanisms: Formal alignment with international standards (WHO [3], UNESCO [27], EU AI Act [4]) for high-risk medical AI; regulatory oversight requiring pre-market conformity assessments for bias and transparency; national ethical guidelines with enforcement mechanisms, not voluntary codes.

Justification from WHO Europe [7]: Only 11% of EU Member States have health-specific AI strategies. Cross-sectoral strategies, while common (85%), do not adequately address clinical risks. System-level mandates are necessary to ensure health-specific governance.

Justification from Abu Dhabi [8]: The Emirati Genome Programme and PHI platform operate under sovereign data governance, but no external regulatory enforcement is described. This framework adds this layer.

6.2 Organisational-Level Governance (Hospitals & Institutions)

Mechanisms: AI ethics committee including clinicians, data scientists, legal experts, and patient representatives; structured bias auditing before deployment and periodically thereafter, stratified by age, sex, race/ethnicity, and socioeconomic status; data governance policies specifying provenance, representativeness, and consent for training data; clinician AI literacy training covering basic AI operation, limitations, and override protocols.

Justification from WHO Europe [7]: Only 26% offer in-service AI training; only 48% have created new AI/data science roles; only 24% have post-market monitoring agencies. Organisational mechanisms are urgently needed.

Justification from Abu Dhabi [8]: Malaffi and Shafafiya provide audit trails, but explicit bias audits are not mandated. The proposed framework makes this a requirement.

6.3 Clinical-Level Governance (Practice & Decision-Making)

Mechanisms: Human-in-the-loop decision models with clear documentation of whether an AI recommendation was followed or overridden; transparent patient communication protocols for discussing AI involvement, including a plain-language explanation of benefits, risks, and the option to decline AI-assisted decisions; integration of AI into shared decision-making as a decision aid, not a substitute for clinical dialogue.

Justification from WHO Europe [7]: Only 18% of EU Member States consult the broader public on AI governance; only 36% consult patient associations. Clinical-level communication protocols are absent in most countries.

Justification from Abu Dhabi [8]: The UMOC command centre uses AI for emergency coordination, but override protocols are not specified. The Sahatna app provides patient-facing insights, but consent scripts are not detailed. This framework fills these gaps.

7. Implications for Healthcare Management

 

 

Table 6. Summary of Implications for Different Stakeholders

Stakeholder	Key Implications	Priority Actions
Healthcare Leaders	AI governance must be embedded into strategic planning, not treated as an add-on	Establish AI ethics committees; mandate bias auditing; require override documentation
Policy Makers	Harmonised regulatory frameworks with enforcement mechanisms are urgently needed	Move from voluntary codes to enforceable regulations; mandate post-market surveillance
Medical Educators	AI ethics and governance must be integrated into curricula at all levels	Develop tiered training programmes on AI fundamentals, ethics, and clinical integration
AI Developers	Products must be designed for auditability, explainability, and human oversight	Build in bias testing, transparency documentation, and override interfaces
Patient Advocates	Patient and public engagement must be mandated, not optional	Require public consultations and patient representation on ethics committees
Researchers	Empirical studies of governance framework effectiveness are needed	Test framework feasibility, bias reduction, and impact on clinician behaviour

Source: Author's original synthesis of implications derived from: (1) governance gaps identified in the WHO Europe [7] survey (Table 1); (2) WEF Abu Dhabi 2026 gaps (Table 3); (3) operationalised governance gap analysis (Table 4); and (4) the proposed three-level governance framework (Table 5).

 

7.1 For Healthcare Leaders

The WHO Europe [7] data show that only 24% of countries have agencies responsible for post-market monitoring of AI systems. Healthcare leaders should establish internal AI ethics committees, mandate pre-deployment and annual bias audits, and require documentation of clinician overrides. Proactive governance reduces liability and improves patient trust.

7.2 For Policy Makers

The EU AI Act [4] provides a regulatory foundation, but WHO Europe [7] data show that national implementation lags: only 11% have health-specific strategies, and only 7% have ethical guidelines. Policy makers must move from voluntary codes to enforceable regulations with specific requirements for bias auditing, transparency, and liability.

7.3 For Medical Education

Only 26% of EU Member States offer in-service AI training for health professionals [7]. Medical curricula must integrate AI ethics, data governance, and human-in-the-loop protocols. Training should be tiered: basic AI literacy for all clinicians, advanced skills for AI leads.

7.4 For Future Research

Empirical studies are needed to test the proposed framework in real-world settings. Key research questions include: Does mandatory bias auditing reduce measurable disparities in AI diagnostic accuracy? Do clinician override protocols reduce automation bias? Do standardised patient communication scripts improve informed consent and trust? How does the framework perform in low-resource settings compared to high-resource settings like Abu Dhabi?

8. Conclusion

 

Artificial Intelligence holds transformative potential for clinical decision-making, but its ethical and governance challenges cannot be overlooked. This systematic review demonstrates that ethical risks - algorithmic bias, diminished physician autonomy, compromised informed consent, accountability gaps, and changes to the patient–provider relationship - are systemic and interconnected. Moreover, current ethical frameworks remain largely conceptual; a significant implementation gap exists.

Cross-country evidence from the WHO European Region [7] quantifies this gap: only 11% of EU Member States have health-specific AI strategies, 7% have health-specific ethical guidelines, and 11% have dedicated liability frameworks - this despite the EU AI Act having taken effect. Only 26% offer in-service AI training, and only 18% consult the broader public on AI governance.

The World Economic Forum [8] white paper on Abu Dhabi's intelligent health system shows what is possible when data infrastructure, political will, and investment converge. The emirate has reduced emergency response times, lowered cancer screening ages, and integrated over 100,000 data streams. Yet even this advanced system, precisely because it is a pioneer, reveals opportunities for further governance refinement, including explicit mechanisms for clinician override protocols, patient consent scripts, mandatory bias audits, and clear liability pathways. What Abu Dhabi has achieved in data integration and clinical outcomes provides a strong foundation; what this framework proposes is the natural next step in its ongoing journey toward fully ethical, patient-centred AI governance.

As operationalised in Table 4, the gap is not in technical capability or ethical acceptance, but in what is mandated, operationalised, and enforced. The governance gap exists because mandatory mechanisms for pre-deployment audits, annual re-audits, demographic stratification, public reporting, and enforcement are absent across most systems, including advanced exemplars.

The proposed three-level governance framework provides a practical pathway for implementing ethical AI in healthcare. By specifying concrete mechanisms at the system, organisational, and clinical levels - AI ethics committees, mandatory bias auditing, human-in-the-loop protocols, and transparent patient communication protocols - it enables healthcare organisations to move from abstract principles to accountable practice.

Ultimately, the success of AI in medicine depends not only on technological advancement but on our collective ability to ensure that AI remains human-centred, transparent, and accountable. As Abu Dhabi demonstrates, intelligent health systems are achievable. As the WHO Europe data demonstrate, most countries are not yet there. The next frontier is embedding ethical governance as deeply as data infrastructure - ensuring that personalisation at scale does not come at the cost of equity, transparency, or trust.

 

Funding

This research received no external funding.

Informed Consent Statement

Not applicable

Conflict of Interest

None declared