Evaluation Frameworks for Clinical AI Incorporating Validation Strategies, Real-World Applicability, and Ethical Principles: Scoping Review

Introduction

AI has evolved substantially since its conceptualization in the mid-20th century, when early research explored the possibility of machines emulating human cognitive functions. Driven by advances in computational power, algorithm development, and access to large-scale datasets, AI has expanded from theoretical domains to widespread applications, including health care. Current uses include medical image analysis, disease risk prediction, drug discovery, and treatment personalization [1-3].

In recent years, the number of AI applications proposed for health care has increased rapidly, with hundreds of models developed for diagnostic support, risk prediction, and workflow optimization. However, only a limited proportion of these models have undergone rigorous external validation or prospective clinical evaluation, raising concerns about their readiness for real-world clinical implementation [4-6].

Despite these advances, several challenges continue to limit the integration of AI into clinical practice [7]. These include variability in data quality, limited external validation, lack of transparency in algorithm development, and insufficient assessment of clinical impact in real-world settings [8]. Ensuring accuracy, safety, and reproducibility, while addressing ethical and regulatory requirements, remains essential. The absence of standardized validation procedures and reporting practices represents a major barrier to reliable clinical adoption.

In response, multiple guidelines and evaluation frameworks have been developed to improve the quality, transparency, and clinical applicability of AI models [9]. This lack of standardized evaluation approaches has prompted increasing concern among clinicians, regulators, and researchers regarding the reproducibility, safety, and transparency of AI-based clinical tools [10].

Notable initiatives include the TRIPOD+AI (Transparent Reporting of Multivariable Prediction Models with Artificial Intelligence) [11], the CONSORT-AI (Consolidated Standards of Reporting Trials for Artificial Intelligence) [12], and the CLAIM (Checklist for Artificial Intelligence in Medical Imaging) [13]. More recently, additional initiatives such as STARD-AI (Standards for Reporting Diagnostic Accuracy) [14] and PRISMA-AI (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Artificial Intelligence) [15] have further expanded reporting standards for diagnostic accuracy studies and systematic reviews involving AI technologies. These frameworks provide structured approaches for assessing data quality, methodological rigor, performance metrics, and clinical utility [16].

However, substantial heterogeneity persists across these initiatives, particularly in methodological scope, validation strategies, and the integration of ethical considerations. This variability makes it difficult for researchers and health care systems to identify appropriate approaches for evaluating AI systems intended for clinical use [17,18].

The objective of this scoping review was to systematically map, characterize, and critically analyze existing evaluation and reporting frameworks for clinical AI. Specifically, this review aimed to examine how these frameworks conceptualize and operationalize: (1) methodological rigor and validation strategies, including internal and external validation; (2) claims related to clinical applicability and real-world use; and (3) the integration of ethical principles, with particular emphasis on alignment with the 10 ethical domains proposed by the United Nations Educational, Scientific and Cultural Organization (UNESCO).

Methods

Protocol and Registration

We conducted a scoping review incorporating a quantitative documentary analysis to examine evaluation frameworks for AI models in clinical practice. The methodology was guided by the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) [19], the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 for Abstracts Checklist [20], and was registered in PROSPERO (CRD420251019640).

The process involved the following two main stages: (1) a systematic search and selection of information sources and (2) data extraction and comparative analysis across predefined thematic domains. These domains encompassed (1) the general characteristics of each AI evaluation framework, (2) validation aspects (eg, performance metrics and generalizability), and (3) the integration of ethical principles.

Eligibility Criteria

All searches covered literature up to February 28, 2026, with no start date limit and no language or publication status restrictions applied.

We included documents that proposed or described an evaluation or reporting framework for clinical AI, encompassing the following source types:

• Peer-reviewed scientific papers: publications in scientific journals presenting frameworks or guidelines for AI evaluation.
• Gray literature documents: documents such as white papers, technical reports, or policy guidelines not formally published in peer-reviewed journals.
• Official websites of relevant organizations: frameworks or directives available on official websites of relevant health, research, or regulatory organizations.
• Introductory series or conceptual papers without real-world application or empirical evaluation.
• Comparative or descriptive analyses of existing AI checklists.
• Documents from out-of-scope domains, including expert conference proceedings
• Letters to the editor, commentaries, or editorial responses lacking original data
• Other reasons

Publications were excluded if they did not present a tangible AI evaluation framework or were outside the scope of our review. Specifically, we excluded methodological or development protocols without validation results or empirical testing of AI models; introductory series or conceptual papers without real-world application or empirical evaluation; comparative or descriptive analyses of existing AI checklists; documents from out-of-scope domains, including expert conference proceedings; letters to the editor, commentaries, or editorial responses lacking original data; and other reasons. Duplicate frameworks reported in multiple sources were counted only once.

Information Sources

Comprehensive searches were performed in six electronic databases: PubMed (MEDLINE), BVS, Embase, EBSCOhost (including CINAHL and others), Sage Journals, and ProQuest Central. The process was conducted independently by DCLM and MH-P.

In addition, we searched pertinent gray literature sources. The EQUATOR (Enhancing the Quality and Transparency of Health Research) Network was used to identify relevant reporting guidelines for AI, and we screened the official websites of major organizations and consortia involved in clinical AI for any published frameworks [21]. This search was conducted by NMA and DCLM.

Search

The search strategy combined controlled vocabulary terms (eg, MeSH and Emtree headings) with free-text keywords, ensuring comprehensive coverage of the following three concept areas: (1) AI and machine learning in health care, (2) model validation and performance evaluation, and (3) ethical principles and frameworks. An example PubMed search strategy is provided below:

(“Artificial Intelligence”[Mesh] OR “Artificial Intelligence”[tiab] OR “Machine Learning”[Mesh] OR “Machine Learning”[tiab] OR “Deep Learning”[tiab] OR “Neural Networks, Computer”[Mesh] OR “Neural Network*”[tiab] OR “Natural Language Processing”[Mesh] OR “Natural Language Processing”[tiab] OR “Generative AI”[tiab] OR “Explainable AI”[tiab]) AND (“Validation Studies as Topic”[Mesh] OR “Validation”[tiab] OR “External Validation”[tiab] OR “Model Validation”[tiab] OR “Generalizability”[tiab] OR “Calibration”[tiab] OR “Performance Metrics”[tiab] OR “Reproducibility of Results”[Mesh] OR “Overfitting”[tiab]) AND (“Decision Support Systems, Clinical”[Mesh] OR “Clinical Decision Support”[tiab] OR “Clinical Decision-Making”[Mesh] OR “Predictive Models”[Mesh] OR “Prediction Model*”[tiab] OR “Prognostic Model*”[tiab] OR “Risk Prediction”[tiab] OR “Diagnostic Accuracy”[tiab] OR “Computer-Assisted Diagnosis”[Mesh]) AND (“Ethics, Medical”[Mesh] OR “Bioethics”[Mesh] OR “Ethic*”[tiab] OR “Equity”[tiab] OR “Transparency”[tiab] OR “Accountability”[tiab] OR “Informed Consent”[tiab] OR “Patient Autonomy”[tiab] OR “Privacy”[tiab] OR “Confidentiality”[tiab] OR “Responsible AI”[tiab] OR “Evaluation Framework*”[tiab] OR “Reporting Guideline*”[tiab] OR “CONSORT-AI”[tiab] OR “SPIRIT-AI”[tiab] OR “TRIPOD-AI”[tiab] OR “STARD-AI”[tiab] OR “CLAIM”[tiab]).

Equivalent search queries were adapted for each database using the respective controlled vocabulary and syntax (refer to Table S1 in Multimedia Appendix 1 for the full search strategies).

Selection of Sources of Evidence

The search strategies were reviewed by the study coauthors. All references retrieved were exported into Microsoft Excel, which was used for reference management, deduplication, and subsequent screening steps. First, 2 reviewers (ACHA and NMA) independently screened the titles and abstracts of all unique records against the eligibility criteria. Next, full-text papers or documents were obtained and assessed for inclusion by the same 2 reviewers working in duplicate. Any discrepancies or uncertainties in study selection were resolved through discussion; if consensus could not be reached, a third reviewer served as an adjudicator.

Data Charting Process

A standardized data charting form was used to extract relevant information from each included source, structured according to the 3 key domains of interest. For each included framework, data were charted and organized into comparative tables corresponding to general framework characteristics, validation and performance evaluation aspects, and ethical considerations.

Two reviewers collaboratively performed the data extraction, with cross-verification by the broader team to ensure consistency. Any discrepancies were resolved through discussion, and when consensus could not be reached, a third reviewer adjudicated (DCLM).

General Characteristics

General characteristics extracted and compared across frameworks included development team composition, disciplinary scope, methodological approach, structural domains, extensions, and stated purpose. Development team composition was categorized as unitary, binary, ternary, or quaternary, including the involvement of one, two, three, or four or more institutions, respectively. Disciplinary scope was defined as interdisciplinary if the team included members from different academic fields, and transdisciplinary if it also incorporated nonacademic stakeholders. Structural domains were grouped into IMRD (Introduction, Methods, Results, and Discussion) and non-IMRD formats. Methodological approach, stated purpose, and extensions were recorded exactly as reported in each evaluation framework.

Validation and Performance Aspects

We extracted details to understand each framework’s orientation toward clinical use and its recommended validation stringency. Each framework’s intended purpose or primary use case was noted and categorized as either research-focused (primarily for studies developing or reporting AI models) or clinical utility-focused (intended to guide evaluation for real-world implementation). For clinically oriented frameworks, we further noted the specified application domain (eg, diagnostic vs prognostic models) and any criteria used to define “clinical utility” [9].

To appraise methodological rigor, we collected information on the following three key validation parameters defined a priori: technical performance, clinical performance, and generalizability [22]. Technical performance (internal validity) measured using metrics such as sensitivity, specificity, area under the curve (AUC) or receiver operating characteristic (ROC), and free-response ROC curve (FROC). ROC and FROC were classified as internal validation parameters, as they are primarily influenced by threshold effects rather than disease prevalence [23,24]. Clinical performance (external validity), assessed using predictive values (positive and negative) and calibration accuracy in AI models. These metrics are influenced by disease prevalence and therefore better reflect real-world performance [22]. Generalizability, defined as the framework’s ability to account for limitations related to uncontrolled overfitting. Overfitting occurs when an algorithm is trained and tested on the same dataset used for development or when internal data splits are used instead of independent or external datasets for validation [25].

• Technical performance (internal validity): measured using metrics such as sensitivity, specificity, AUC or ROC, and FROC. ROC and FROC were classified as internal validation parameters, as they are primarily influenced by threshold effects rather than disease prevalence [23,24].

For this review, evaluation frameworks were classified a priori into three levels of methodological rigor with respect to their assessment of clinical utility and generalizability:

• Complete evaluation: defined as frameworks that require validation using independent external datasets distinct from those used for model training, while also considering the epidemiological study design (eg, diagnostic cohort studies or case-control designs). This approach ensures both internal validity (the model performs adequately in its development sample) and external validity (the model retains performance across real-world clinical settings). Frameworks in this category provide the strongest evidence of clinical reliability, as they minimize overfitting and explicitly account for study design quality.
• Partial evaluation: defined as frameworks that acknowledge the importance of external validation datasets but do not assess the methodological quality or epidemiological design of those datasets. Although this category represents a step toward generalizability, it does not guarantee that external data are sufficiently robust or representative. Consequently, such frameworks provide only limited evidence to support clinical implementation, as risks of selection bias and prevalence-related distortions remain.
• Inadequate evaluation: defined as frameworks that consider AI models to be generalizable even when validation is based solely on internal datasets, typically through data splitting or cross-validation within the same population. This approach fails to ensure external validity and is highly susceptible to overfitting, as reported performance primarily reflects adaptation to the original dataset. Frameworks in this category should not be regarded as sufficient to support clinical applicability; rather, they are relevant only for the developmental phase of model assessment.

Synthesis of Results

We synthesized the charted data through a descriptive comparative approach, emphasizing both the distribution of characteristics across frameworks and the identification of patterns or gaps in current practice. Extracted information was tabulated and compared across the three thematic domains to facilitate a side-by-side analysis of frameworks (eg, allowing readers to contrast different frameworks’ requirements for validation or ethical compliance).

A dot plot was used as a graph to present the gap map, following the approach suggested by Nyanchoka et al [26]. The evaluation of gaps was performed across 3 predefined domains. For each domain, frameworks were assessed according to specific criteria, and scores were assigned to indicate the presence or absence of gaps.

In Domain 1 “general characteristics,” frameworks were classified as having no gap (1 point) when they incorporated quaternary teams and a transdisciplinary scope. Frameworks that only included an interdisciplinary scope or other team compositions were considered to have a gap (0 points).

In Domain 2 “validation and performance aspects,” frameworks received a score of no gap (1 point) when they addressed clinical utility and provided comprehensive validation (complete generalizability) and appropriate performance considerations. Frameworks that did not explicitly include these elements were classified as having a gap (0 points).

In Domain 3 “ethics,” frameworks were considered to have no gap (1 point) when they explicitly incorporated direct and indirect principles and did not present missing ethical components (direct + incorrect and no missing components). Frameworks lacking these elements were categorized as having a gap (0 points).

This scoring approach allowed the identification of gaps across domains and facilitated the graphical representation of the distribution of strengths and weaknesses among the evaluated frameworks.

We did not perform a formal quantitative meta-analysis given the qualitative nature of the data; however, we carried out simple quantitative aggregations (counts and percentages) for certain features to summarize prevailing trends (reported in the Results section).

Protocol Deviations

Deviations from the initial protocol included updating the literature search at a later date than originally planned to ensure the currency of the evidence base, adding a gap map to enhance the presentation of results, incorporating PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature search extension) reporting elements to improve transparency of the search process, refining the approach to results synthesis, and expanding the number of databases searched. These modifications were made to strengthen the completeness, transparency, and interpretability of the review and did not change its primary objective.

Ethical Considerations

The study was approved by the Research Ethics Committee of the Universidad Cooperativa de Colombia (protocol INV3688). The ethical evaluation of the included frameworks followed the UNESCO ethical principles for AI [27]. An ethical scoring matrix (Table S2 in Multimedia Appendix 1 [13-15,17,28-69]) was used to assess each framework across the 10 UNESCO domains, with predefined criteria indicating whether each principle was explicitly addressed (fully covered with clear guidance), partially aligned (mentioned or implied but not fully detailed), or absent. Each principle (eg, transparency, fairness, privacy, human oversight, and accountability) was evaluated individually using a semiquantitative scoring system. Domain-level scores were then calculated to compare the ethical coverage across frameworks. Two reviewers conducted the assessment independently, resolving discrepancies through consensus. This structured matrix-based approach ensured a systematic and reproducible analysis of the ethical strengths and gaps within current AI evaluation frameworks used in clinical practice.

Results

Study Selection

A total of 3363 records were identified through database and gray literature searches. After removal of duplicates, 391 records were screened at the full-text level. Of these, 345 were excluded for not meeting the eligibility criteria, primarily because they did not propose or describe an evaluation framework for clinical AI or lacked evaluative content. Ultimately, 46 distinct evaluation frameworks were included in the scoping review. Figure 1 presents the PRISMA flow diagram detailing the number of records identified, screened, excluded, and finally included in the review. Table 1 summarizes all included papers and their main characteristics, while the complete list of excluded studies with corresponding reasons is available in Multimedia Appendix 2.

‎

Domain 1: General Characteristics

The included publications, spanning from 2016 to 2025, illustrate a substantial expansion of methodological and reporting initiatives aimed at strengthening the evaluation of AI systems in health care. A majority of frameworks (63%, 29/46) were developed through quaternary, multi-institutional collaborations, and 32.6% (15/46) involved transdisciplinary participation, including regulators, industry partners, policymakers, or patient representatives.

With respect to methodological development, more than half of the frameworks were created using formal consensus processes, such as Delphi rounds, modified Delphi methods, consensus meetings, or iterative multistakeholder refinement. Among these, several represent EQUATOR-endorsed extensions (eg, CONSORT-AI, SPIRIT-AI [Standard Protocol Items – Recommendations for Interventional Trials – Artificial Intelligence Extension], Developmental and Evaluation Checklist for Decision Support Systems Driven by Artificial Intelligence [DECIDE-AI], CLAIM 2024, TRIPOD-Cluster [TRIPOD Extension for Clustered Data], TRIPOD + AI, CHEERS-AI [Consolidated Health Economic Evaluation Reporting Standards for Interventions That Use Artificial Intelligence], and STARD-AI). An additional 13% (6/46) were derived from systematic or scoping reviews. Only 2 frameworks (4.3%), APPRAISE-AI (Assessment of Predictive Performance and Bias in Artificial Intelligence Studies) and the AI-CDSS success framework, underwent empirical validation procedures, including reliability testing, construct validity assessment, or psychometric modeling.

Structural formats also varied: 65% (30/46) of the documents followed an IMRD-compliant structure. Across all frameworks, recurring domains included dataset description, model development procedures, performance metrics, internal and external validation strategies, reproducibility requirements, and overall reporting transparency. Several frameworks were specialty-specific, addressing fields such as dermatology, dentistry, cardiology, urology, or interventional radiology.

Importantly, 52.2% (24/46) of the documents corresponded to official extensions of established reporting guidelines (PRISMA or EQUATOR) or to independently implemented frameworks aligned with these standards.

Taken together, the included documents reveal a rapidly expanding and increasingly structured landscape of AI evaluation frameworks, characterized by high levels of interdisciplinarity, growing methodological convergence, and a clear movement toward formalized reporting standards designed to enhance transparency, reproducibility, and clinical applicability in AI-based health care research.

Domain 2: Validation and Performance Aspects

Two documents were not included in Table 2 because they addressed exclusively ethical considerations and did not provide evaluative components related to technical performance, clinical effectiveness, or clinical purpose.

The evaluation frameworks revealed substantial heterogeneity in how studies articulate their objectives, define clinical purpose, and report technical or clinical performance parameters. Most frameworks were investigational in nature (approximately 88%, 39/44), whereas only 4 frameworks explicitly stated a focus on assessing clinical utility (eg, Reps et al [30], Hernandez-Boussard et al [32], Ji et al [40], and Olczak et al [41]). This distribution indicates that most published frameworks remain exploratory rather than implementation-oriented.

Across studies, technical performance reporting was inconsistent. Only 31.8% (14/44) provided at least 1 performance metric. The most frequently reported indicators were AUC or ROC values, present in 12/44 studies, followed by sensitivity and specificity. Less commonly, frameworks used positive predictive values and negative predictive values, accuracy, or segmentation metrics such as the F₁-score or Dice coefficient (eg, Norgeot et al [35] and Olczak et al [41]). A small subset incorporated measures of calibration accuracy, number needed to treat, or task-specific metrics (eg, 30-day mortality in Kalaycıoğlu et al [66]).

Reporting of clinical effectiveness metrics was even more limited. Only 7 (15.9%) studies explicitly referenced indicators of clinical performance, most commonly predictive values, accuracy, or calibration. The overwhelming majority did not report any clinical effectiveness parameters, suggesting that most frameworks remain focused on model development and internal performance rather than real-world clinical applicability.

When classifying the rigor of evaluation (technical performance), only 5 (11.4%) frameworks were categorized as providing a complete evaluation, integrating internal and external validation components (eg, Luo et al [28], Hernandez-Boussard et al [32], Mongan et al [34], Norgeot et al [35], Kwong et al 2021, and Debray et al [51]). An additional 5 (11.4%) studies were judged as incomplete, and 3 studies were rated as incorrect (“wrong”) due to methodological inconsistencies or misalignment between reported metrics and intended clinical purpose.

Regarding clinical purpose, approximately one-third (9/44) of the frameworks focused on diagnostic applications, while 7 of 44 addressed prediction. Another subset (7/44) incorporated both diagnostic and prognostic applications.

Domain 3: Ethical Alignment With UNESCO Principles

Analysis of the ethical dimensions of the included frameworks revealed substantial variability in their alignment with the 10 UNESCO AI ethical principles. Overall compliance was modest, with only a small group of documents achieving high scores (≥80%), including works by Floridi et al [29], Shen et al [48], Masters and Salcedo [62], Ning et al [17], and the PAHO [67], each demonstrating broad and explicit ethical integration. In contrast, four documents (Luo et al [28], Stevens et al [37], Abdulazeem et al [50], and Debray et al [51]) showed no or minimal ethical reporting, scoring below 10%. Across principles, the highest levels of adherence were observed for awareness and education (P7, 71.1%), transparency and explainability (P9, 70%), and proportionality and safety (P2, 63.3%), indicating a stronger emphasis on communicability, system clarity, and risk mitigation. Conversely, human oversight (P4, 24.4%) and adaptive, multistakeholder governance (P8, 33.3%) were the least addressed, reflecting limited incorporation of human-in-the-loop safeguards and participatory governance structures. Only a minority of frameworks exhibited balanced coverage across all principles, revealing a fragmented ethical landscape where most documents address isolated ethical components rather than offering comprehensive, principle-wide guidance (Figure 2 [13-15,17,28-69]).

‎

The gap map presented in Figure 3 [13-15,17,28-69] shows that most existing AI evaluation frameworks primarily emphasize methodological and technical aspects of model performance. In contrast, comparatively less attention is given to clinical applicability and ethical considerations. The distribution of data points across the 3 evaluated domains suggests that these frameworks tend to prioritize reporting transparency and reproducibility, while aspects related to clinical utility, external validation, and ethical safeguards for real-world implementation remain insufficiently addressed.

Across the three domains evaluated (general characteristics, validations and performance aspects, and ethics), most frameworks presented substantial gaps. Specifically, 63% (29/46) of frameworks showed gaps across all 3 domains. In contrast, 34.8% (16/46) of the frameworks fully complied with the criteria in at least 1 domain. Among these, 3 frameworks met the criteria for the ethics domain (Floridi et al [29], Masters and Salcedo [62], and Ning et al [17]), while 13 complied with the general characteristics domain. Only 2.2% (1/46) of the frameworks met the high-quality criteria in 2 domains (general characteristics and ethics). Importantly, none of the evaluated frameworks met the high-quality criteria for the validation and performance aspects domain, and no framework fulfilled the criteria across all 3 domains simultaneously.

‎

Discussion

This scoping review identified a fragmented landscape of evaluation frameworks for clinical AI, characterized by heterogeneous methodological approaches, limited integration of external validation aligned with intended clinical use, and inconsistent incorporation of ethical principles. These findings directly address the study objectives by highlighting gaps in methodological rigor, real-world applicability, and ethical alignment. Ethical integration was inconsistently addressed, with most frameworks covering only a limited subset of UNESCO ethical domains.

These findings can be better understood in the context of the rapid expansion of methodological and reporting frameworks for clinical AI [6]. This pattern is consistent with the orientation of several influential frameworks included in this review, many of which were developed primarily to improve transparency, reproducibility, and completeness of reporting, while providing more limited coverage of implementation-related issues [72]. Collectively, this proliferation of frameworks reflects a broad consensus regarding the need to promote transparency, rigor, and reproducibility in medical AI research [29,34,49]. Minimum Information for Medical AI Reporting (MINIMAR), for example, was proposed to improve transparent reporting of the design, development, evaluation, and validation of medical AI models, with explicit emphasis on replication, external validation, and the identification of potential biases and unintended consequences [32]. Likewise, Minimum Information about Clinical Artificial Intelligence Modeling (MI-CLAIM) was designed to enable assessment of clinical impact, fairness, and bias while facilitating replication of the technical design process of clinical AI studies [35]. In the same direction, Klement and El Emam [53] showed that existing high-quality guidelines did not individually provide complete coverage, supporting the fragmented landscape identified in our gap map. However, our findings suggest that this expansion has occurred without corresponding consolidation of validation standards or clinical applicability.

The predominance of technical and reporting-oriented domains over broader clinical dimensions in our review also aligns with previous literature showing that strong model performance does not necessarily translate into clinical benefit. Lu et al [47] found inconsistent adherence to reporting recommendations in deployed clinical prediction models, particularly for reliability-related items such as external validation, uncertainty, monitoring, fairness, and transparency. This broader concern is also reflected in DECIDE-AI, which notes that strong performance in preclinical studies has not yet been matched by high-quality evidence of improved clinician performance or patient outcomes in clinical settings [49]. Taken together, these findings suggest that current evaluation of clinical AI remains more mature at the level of methodological description than at the level of demonstrated clinical usefulness [46,48].

A related implication is that methodological rigor in clinical AI should not be interpreted narrowly as internal validation or reporting compliance. SPIRIT-AI and CONSORT-AI both emphasize that AI interventions should be described in relation to their intended use, intended users, integration into the clinical pathway, required expertise, handling of input data, output interpretation, and the way outputs contribute to downstream clinical decision-making [31,33]. CONSORT-AI also requires reporting of performance-error analysis and highlights the importance of documenting access to the intervention or its code, while SPIRIT-AI recommends explicit planning for performance-error identification and description of implementation requirements in the trial setting [31,33]. Together, these frameworks support the view that robust evaluation of clinical AI requires attention not only to methodological transparency but also to usability, context of use, and conditions for safe implementation.

Our finding that real-world applicability was less consistently represented is also supported by frameworks explicitly designed to address translational stages of evaluation [73,74]. DECIDE-AI was developed as a stage-specific reporting guideline for early, small-scale, live clinical evaluation and focuses on proof of clinical utility, safety, human factors, and preparation for larger-scale studies [49]. Its rationale explicitly refers to the gap between mathematical performance and clinical utility and to the need to address challenges such as dataset shift, user variability, and implementation in live clinical settings [49]. More recently, Labkoff et al [61] argued that responsible AI-enabled clinical decision support requires a more comprehensive framework spanning model training, explainability, validation, certification, monitoring, continuous evaluation, privacy, fairness, and regulatory oversight. These proposals are consistent with the pattern observed in our review: implementation-relevant domains are increasingly recognized in the literature, but they remain unevenly operationalized across available frameworks.

The limited number of transdisciplinary studies in our review is also noteworthy. Some of the more mature frameworks were developed through broad multistakeholder processes. DECIDE-AI was produced through an international consensus process involving multiple stakeholder groups [49] and Labkoff et al [61] similarly reported a consensus process including clinicians, software developers, academics, ethicists, attorneys, policy experts, scientists, and patients. However, our mapping suggests that such breadth of stakeholder participation is not yet typical across the field. This matters because clinical AI is not only a technical intervention, but also a clinical, organizational, ethical, and regulatory one. Frameworks developed from narrower perspectives may therefore be less able to capture the full conditions needed for trustworthy implementation [75].

The ethical findings in this review are likewise consistent with the literature included in the scoping review. Floridi et al [29] synthesized major AI ethics initiatives around beneficence, nonmaleficence, autonomy, justice, and explicability, arguing that explicability combines intelligibility and accountability and is necessary to make the other principles actionable. Their discussion of justice also extends beyond discrimination to shared benefit, equal access, and protection of social structures such as health care systems [29]. In more applied health care literature, Shen et al [48] proposed an ethics checklist structured around informed consent, equity, diversity and access, privacy and partnerships, regulation and law, return of results, and duty to warn and report [48]. Ning et al [17] similarly proposed the TREGAI (Transparent Reporting of Ethics for Generative Artificial Intelligence) checklist for generative AI in health care, derived from a scoping review and intended to operationalize ethical evaluation within practical guidance. These studies support our interpretation that ethical concerns are clearly present in the literature, but are not yet incorporated with comparable depth or consistency across clinical AI evaluation frameworks. This limitation is particularly relevant for clinical implementation, as insufficient ethical integration may undermine patient safety, public trust, and regulatory acceptance of AI systems in health care.

A similar pattern can be seen in more implementation-oriented proposals. Bazoukis et al [43] describe a regulatory framework for augmented intelligence in medicine that includes accountability, liability, equity and inclusion, transparency, explainability, education, patient engagement, cybersecurity and privacy, ethics and fairness, as well as safety and postmarket surveillance. Labkoff et al [61] also emphasize transparency, documentation, training, validation, certification, monitoring, safety reporting, fairness, and privacy as necessary elements of responsible AI-enabled clinical decision support. In our review, however, these ethical- and governance-related elements were not uniformly represented across frameworks. This suggests that the current landscape remains fragmented: ethical principles are increasingly acknowledged, but their operational integration into evaluation standards is still incomplete.

Taken together, the three domains examined in this scoping review reveal a recurrent pattern: methodological rigor, clinical applicability, and ethical integration do not consistently converge within individual frameworks. Many frameworks are strong in reporting transparency or technical assessment, but are less explicit regarding external validation, implementation conditions, real-world use, or comprehensive ethical governance. This fragmentation appears to be a structural feature of the current evaluative landscape rather than a weakness of any single framework. It likely reflects the parallel evolution of technical reporting, clinical evaluation, and ethical guidance in AI, which have not yet been fully harmonized within a single evaluative model [29,31-35,43,49,53,61].

Limitations

As with any scoping review, this study has inherent limitations. First, as a scoping review, its purpose was to map and characterize the available literature rather than to compare the superiority or effectiveness of individual frameworks [19,76]. Therefore, the findings should be interpreted as an overview of patterns, emphases, and gaps in the field, not as evidence that one framework performs better than another. Second, the included publications were heterogeneous in scope, purpose, and intended stage of use, ranging from reporting guidelines and consensus statements to conceptual ethics papers and implementation-oriented proposals [29,31,33,34,49,56]. Although this heterogeneity is informative for evidence mapping, it limits direct comparability across sources [15,34,40,45,46].

Third, the gap map required interpretive judgment when classifying whether a framework addressed specific methodological, clinical, or ethical domains, particularly when such issues were mentioned implicitly rather than operationalized explicitly [37,50,51]. For that reason, the categorizations presented here should be understood as a structured synthesis of the literature rather than a definitive ranking. Finally, many of the included publications propose how clinical AI should be evaluated, but do not prospectively test the effect of those frameworks in routine clinical settings [40,54,77]. Accordingly, this review is stronger in identifying what the literature currently prioritizes and omits than in determining how well these frameworks perform in practice [49,61].

Conclusions

In conclusion, this scoping review indicates that current evaluation and reporting frameworks for clinical AI are strongest in technical and reporting-oriented domains, but less consistently address broader validation strategies, real-world applicability, and comprehensive ethical integration. This interpretation is consistent with the included literature, which has emphasized transparency, reproducibility, intended use, validation, fairness, and safety, while also recognizing the persistent gap between algorithm development and responsible implementation in clinical settings.

The broader implication is that advancing clinical AI will require a shift from fragmented, reporting-oriented approaches toward integrated evaluation frameworks that align methodological rigor, real-world validation, and ethical governance. Future frameworks should move beyond checklist-based reporting to support implementation, regulatory decision-making, and continuous monitoring in clinical settings. Without such integration, the translation of AI innovations into routine health care practice will remain limited.