Important Ethical, Technical, and Epidemiological Considerations in an AI Tool Production (ETEPAI): Scoping Review

Introduction

In the rapidly evolving landscape of artificial intelligence (AI) development, the convergence of ethical, technical, and research methods considerations is paramount for multidisciplinary teams aiming to achieve success in the production of AI tools [1]. The impact of AI tools after deployment depends on the intricate interplay between the AI system’s development, the decision-making based on its output, and the capacity of the stakeholders involved to take the necessary subsequent actions. Foreseeing, estimating, and designing the AI tools that meet all the challenges before deployment are essential to bridge the chasm between AI tools development and achievable benefits [2] with acceptable risks [3].

However, there is a proliferation of guidelines, checklists, assessments, frameworks, and recommendations [4,5] which could lead to confusion, inconsistency, and inefficiency in adherence and application [6]. Overwhelmed by the many similar yet different and lengthy referent materials, developers and stakeholders may overlook the most relevant and essential practices, rendering many good referent materials less effective, stalling improvement in the AI tools production process, or risking the perpetuation of the poor conduct and reporting of such studies [7]. These include a lack of data and code availability, an absence of or small human comparator groups, and a shortage of real-world clinical relevance, transparency, and inappropriate conclusions, causing overall high risk of bias [8], leading to more research waste [9-11]. To address these issues, there is a pressing need to integrate the various referent materials into a unified, thorough, concise, and clear approach across different study designs to inform about the best practices and ethical standards that accentuate the clinical relevance of AI tools and systems production that integrate into clinical workflow for sustainability throughout the whole developmental process.

For the purposes of this review, we use the term “artificial intelligence” as a broad umbrella to encompass the spectrum of machine learning (ML) models commonly used for clinical tasks, from traditional regression to deep learning. While the technical complexity of these algorithms varies, the fundamental principles of ethical oversight, rigorous epidemiological validation, and postdeployment governance required for safe clinical integration remain consistent.

Methods

Search Strategy and Selection Criteria

Searches for relevant recommendations were conducted through a multipronged search strategy combining systematic database searches with supplementary methods to capture a broad range of relevant recommendations. First, a systematic search was conducted in key academic databases, including PubMed, Google Scholar, and Semantic Scholar. The search strategy used a combination of keywords and Boolean operators to identify relevant literature. A representative search string was: ("artificial intelligence" OR "machine learning") AND ("guideline" OR "framework" OR "checklist" OR "recommendation" OR "best practice") AND ("healthcare" OR "clinical" OR "medicine"). Second, this initial search was supplemented by several methods. We conducted a targeted review of regulatory documents from prominent health and technology organizations, published policy, and gray literature. We also performed citation searching by manually reviewing the reference lists of highly relevant guidelines and systematic reviews to identify eligible papers. This process was further enriched through consultation with domain experts to ensure seminal works were not overlooked. Additionally, we used AI-powered academic search tools such as Elicit, SciSpace, and Perplexity to identify relevant preprints and newly published papers that may not have been fully indexed in traditional databases. Recommendations that were selected must be developed by groups of experts through a systematic and scientific process, provide explicit, actionable recommendations for the design, development, validation, or deployment of clinical AI tools, and must be published in English.

Data Extraction and Synthesis

The included recommended guidelines [3,12-16], checklists [17-23] and the AIPA (Artificial Intelligence Prediction Algorithm) for medical sectors [24], the Stanford’s FURM (Fair, Useful, the Reliable Artificial Intelligence Model) assessment [2], the SUDO (pseudo-label discrepancy) framework [25], and the medical algorithmic audit framework [26], and the STANDING (Standards for Data Diversity, Inclusivity, and Generalisability) Recommendations [27] are summarized with the checklists of the referenced materials provided via links on the tables in the Multimedia Appendix 1 [3,12,13,16-21,24,26-51], and excluded guidelines were discussed further and described in Multimedia Appendix 2 [52-66]. We assimilated the referenced materials into important ethical, technical, and epidemiological considerations for all involved in artificial intelligence tool production (ETEPAI) as critical pointers according to 4 coherent stages of design, development, deployment, and postdeployment in 3 domains of ethics, technical, and epidemiological principles (Figure 1). While these 4 stages are analogous to a standard software development life cycle to ensure familiarity and ease of integration, the critical pointers within each stage are specifically tailored to the unique challenges of developing safe and effective AI tools for health care.

‎

Important concepts and must-have indicators from the referenced materials are deemed sufficient and included in the list of items in ETEPAI as they represent well the rest of the indicators in their respective material. To prime users at the start of using ETEPAI are probing questions that aim to enhance readiness (Textbox 1). Added to these guides on best practices is a section on the common pitfalls and prevailing challenges to highlight and re-emphasize the importance of due consideration to every step in the AI tools development in order to avoid serious consequences. In this approach, ETEPAI is more inclusive and thorough compared to any of the referenced materials. ETEPAI considerations are comprehensive, complete, compact, and crisp to use to guide high-quality, ethically sound, and clinically relevant AI tools from idea conception to plan execution, and from promoting to monitoring. As a result, we also provided an AI research proposal template in Multimedia Appendix 3. Thus, ETEPAI serves as annotated content pages of books that signpost authors and users to the necessity of each step in the process but would require further reading in order to gain further understanding if this did not happen before. Similarly, readers and users of this ETEPAI are advised to refer to the original references for full explanation (Multimedia Appendix 1), and to other more elaborative literatures [67], on rigorous external model validation [16,68-70] and for clinical trial and economic evaluation [21,23,28], or for additional help in complete reporting guidelines [5,71], in organizational capabilities strengthening for AI adoption [72], in sociotechnical dimensions to consider when integrating AI tools in complex adaptive health care systems [73], and to identify technological solutions for achieving large-scale sustained adoption [74]. The scale and brevity of ETEPAI may be sufficient for trained personnel and serve as a sure sign to unmotivated stakeholders, as a standard evaluation and as materials to equip organizations to be AI-capable [72]. Further elaboration on its uses is given below.

ETEPAI’s product-centric approach could effectively align with European Union (EU)–defined trustworthiness standards [29,75] by emphasizing safety, robustness, and ethical considerations at every stage of the AI life cycle, from design through deployment. With its focus on rigorous processes, techniques, and verifiable methods, ETEPAI promotes comprehensive risk identification and mitigation strategies that address both product-level and system-level impacts on users and society. This focus directly supports compliance with the EU’s safety, health, and fundamental rights protections, ensuring AI systems are both reliable and ethically responsible. While ETEPAI aligns well with the principles of the EU AI Act, its fundamental focus on data privacy, responsible data handling, and accountability ensures its principles are compatible with and supportive of other major regulatory frameworks, such as HIPAA (Health Insurance Portability and Accountability Act) in the United States [76].

Results

Critical Pointers and Domains

Figure 2 shows the flow diagram [77]. Figure 1 illustrates the key stages of AI tools from development to postdeployment. Textbox 1 lists probing questions in the 4 stages of design, development, deployment, and postdeployment stages according to their relevance and importance when thinking about having an AI tool in a health care setting. These would alert and engage users’ attention to the scopes and contexts of ETEPAI. Similarly, Table 1 provides critical pointers in AI tools production arranged in 3 domains of ethics, technical, and epidemiological principles under the same 4 stages. Although the probing questions and the critical pointers are grouped into different stages and domains, they are all to be considered when any AI tool is being conceived by an individual or discussed by a team of people, which should either result in a comprehensive research project proposal or abandonment if it is found to be unsound or a nonfeasible idea. Please see Checklist 1 for PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) documentation.

‎

ETEPAI Tool Production and Use

The designing stage is aligned with the epidemiologic research methodologies comprising the theoretical design, data collection design, and statistical analysis design [80,81]. The theoretical design involves identifying the real need for the AI tool, and defining the research question by having a clear target population, the exposure (or independent variable) of interest or interventional program, context or control, and the outcome measured at specified time point and setting (the PE/ICOTS [Population or Participants, Exposure/Intervention, Comparator, Outcomes, Timing, and Setting] acronym) [35,80]. The data collection is usually on already available datasets to be further selected and cleansed according to eligibility criteria. Prospective sampling is instructed by some guidelines. The statistical analysis includes the AI-specific models’ development or selection, evaluation, and validation procedures, estimating the best-performing AI-based prediction model. This is often summarized in terms of calibration and discrimination, without overoptimism, and with internal (cross-validation), external validation, decision curve analysis, (early) health technology assessments, and impact studies (such as via randomized clinical trials) [1]. The development stage is the conduct of the project according to the published and openly available project or research proposal. This allows public scrutiny of the project conduct from the proposed designs. Therefore, justification as a clear explanation must be given for every deviation in the actual conduct of the project from the proposal. The deployment and postdeployment stages are similar to the implementation research [82,83], where the tested AI tools are used in the routine practice. Deployment cost and organizational effort required to integrate the AI systems into a clinical workflow should be estimated to facilitate sustainability [78]. Most pointers in Table 1 have explanatory footnotes with further explanations in the Multimedia Appendix 4, and informative referenced materials given in parentheses. While the ethics domain has its referenced materials in Table S1 in Multimedia Appendix 1. This keeps the table within a readable limit and with adequate explanatory reference information. The explanatory footnotes are taken from the reference materials that are combined for maximal practical and educational clarity.

Common Pitfalls and Prevailing Challenges

Afterthoughts and neglecting the aforementioned crucial considerations for AI tools development would inevitably cause serious problems and costly consequences. These pitfalls include inadequate data quality, biased algorithms, and misaligned AI capabilities that often emerge despite rigorous planning [84]. To re-emphasize the importance of due consideration of ETEPAI at every step throughout the development process to mitigate these risks [84], this section shares some potential pitfalls and appropriate preventive and corrective measures (Table 2) that can lead to models that perform effectively in real-world settings. These notable pitfalls and veracious challenges during AI tools development may vary according to different tools and settings. Further guides on best practices to remedy pitfalls, challenges, errors, and biases are available elsewhere [85,86].

Common Pitfalls to Caution

In the rapidly evolving field of AI, developers and data scientists face a myriad of challenges that can significantly impact the effectiveness and reliability of their models. While much emphasis is placed on the theoretical foundations and potential applications of AI, it is equally important to be aware of the common pitfalls that can undermine these efforts. This section highlights several notable pitfalls that practitioners must navigate to ensure robust and accurate outcomes, from data preparation to model selection and beyond.

One major challenge in AI development is the narrow selection of models and methods [87]. A common pitfall is failing to explore the appropriate model family during model selection, such as using only linear regression models when the data-generating function is nonlinear and discontinuous [88]. Additionally, data scientists or vendors often have strong preferences for a limited set of methods or technologies, even when they are not the most suitable for the task [89]. This narrow approach can significantly hinder model performance and limit the potential of AI applications.

Another common issue is the tendency to rely on single-stage modeling attempts without refinement. Initial models typically require iterative enhancements, as first attempts rarely meet performance goals in complex problems. A single-stage approach lacks the iterative understanding necessary to optimize model interaction with data [90]. Furthermore, ignoring established best methods or misapplying robust methods outside of their intended use can result in underperforming models [91]. Adhering to proven methods and following established guidelines are critical for maintaining model integrity and achieving reliable outcomes.

Insufficient exploration of hyperparameters and data preparation is also a critical issue. Even when the right model families are chosen, inadequate tuning of hyperparameters can lead to suboptimal outcomes [92]. Similarly, improper data preparation, such as neglecting feature construction, selection, normalization, or discretization, can greatly impair a model’s effectiveness [93,94]. Thorough attention to both hyperparameter optimization and data preparation is essential for achieving robust model performance.

Data contamination and bias are additional challenges that can compromise model accuracy. When data processing steps such as normalization or discretization are conducted across both training and test sets, it can introduce bias and lead to inflated performance metrics [95]. This issue is exacerbated when nonrepresentative datasets or unusual populations are used, leading to models that do not generalize well. In some cases, models may produce misleading results by focusing on correlated factors rather than underlying causes, which can result in ineffective or incorrect conclusions.

Overconfidence in models, particularly in the context of overfitting and underfitting, poses a significant risk. Models may appear effective during development but fail in real-world applications due to high-dimensional data, small sample sizes, or overly complex learners [87]. This issue is closely related to the challenge of power-sample calculation in AI, which is more complex than in traditional statistical analysis. Learning curves, which describe the generalization error as a function of sample size, are often unknown, making it difficult to determine the required sample size for reliable model performance.

Addressing common pitfalls is essential for the effective development of AI models, but it is equally important to navigate the more complex challenges that can emerge. By adopting a balanced approach that includes both careful model selection and iterative refinement while also considering algorithmic transparency, regulatory impacts, and complete process reporting, developers can better ensure the reliability and ethical deployment of AI systems. These considerations are crucial for fostering innovation without compromising on responsibility, as AI continues to play an increasingly significant role in various fields.

Veracious Challenges to Circumspect

Navigating the complex landscape of AI development requires a careful balance between transparency, innovation, and responsible implementation. While algorithmic transparency is critical for validating methods, it must be handled with caution to avoid unintended consequences and misuse. Similarly, overengineering and overregulation can stifle innovation, making it essential to strike a balance that promotes both progress and safety. Complete transparency in reporting is also vital, ensuring that all analytical processes are fully disclosed and accessible, enabling thorough evaluations of model performance and potential biases.

Algorithmic transparency is crucial for the validation of methods and tools [96]. However, it also introduces risks, such as the potential for misuse, unintended consequences, or manipulation of models and systems. Black box methods, while generally undesirable when they fail to meet expected safety and performance standards, can be advantageous in scenarios where securing the system against tampering is necessary. Moreover, as argued by some AI literature [97,98], if a well-validated black box model demonstrates a substantial statistical advantage over the best transparent model, it may be both impractical and ethically questionable to disregard its superior performance and use [98]. Navigating these trade-offs requires careful consideration and expertise.

Overengineering and overregulating pose challenges in science and technology [99]. When best practices are enforced in rigid, bureaucratic ways, there is a risk of stifling innovation and slowing progress. Additionally, disguising decision models that guide user actions as mere advisory tools is a persistent issue in health AI [12]. Regulation must address these practices, as they can render regulation ineffective and distort performance and safety requirements during the design of AI systems. Ensuring safety and performance through best practices is essential, but it must be balanced against the risks of delaying the deployment of valuable AI/ML applications in health care and health sciences. Achieving this balance is crucial for fostering both innovation and responsible implementation.

It is essential to practice complete reporting of the full process, especially all the analyses and modeling procedures applied to the data [100,101]. This would enable a thorough evaluation of the robustness and any potential issues of bias. Providing complete access to the algorithms and data, with a detailed disclosure of the entire analytical process, including all model selection and error estimation steps, would definitely help the comprehensive evaluation of model performance and reliability, and its associated aspects, such as overfitting or overconfidence.

Discussion

Principal Findings and Framework Synthesis

This work results in the ETEPAI, which is structured around 4 key stages: design, development, deployment, and postdeployment. It is a wholesome consideration of an AI tool production from designing, development, deployment, and postdeployment stages to be executed in a sound approach and with good adherence to the ethical principles, solid preparation, and execution that meet expected achievable usefulness, information technology feasibility, and crucial technical requirements including deployment strategy, monitoring and evaluation plan, impact assessment, financial projections for sustainability, and take into account what is cohesive to the epidemiologic research methodology of rigorous validation and complete reporting [15]. This is achieved by integrating multiple frameworks and guidelines to ensure a broad and comprehensive approach to AI tool development in health care. This corroborates the declaration of Innsbruck that underscores the evaluation of information systems in health care throughout design and implementation stages [102], and the multistep process design framework of iterative evaluation and system assessment as central to overall development planning [103], and also matches well the key challenges posed by AI in a real-world context of care and services according to the health technology assessment core model [104].

Comparison With Existing Guidelines

Since ETEPAI is essentially a synthesized summary of existing robust recommendations, it is an evidence-based guidance on the tasks it is meant for, which may include using it as an evaluation tool in a systematic review of papers reporting on AI tools development, which would ease the task of using multiple similar checklists [105]. ETEPAI is different from other similar tools, but does not form its referenced base. Multimedia Appendix 2 tabulates the characteristics of these guidelines, checklists, and frameworks. The methodological guidelines tool reviewed the methodologies applied in 134 selected studies and developed its checklist to provide a systematic framework for estimating population-based health indicators using linked data and ML techniques [52], but lacked in certain emphasis, such as on error analysis, guidance on deployment, and postdeployment strategies. Another was based on theoretical knowledge (Clinical Artificial Intelligence Research checklist) [53], one that is discipline-specific that has yet to be developed through the proper processes (Model for Assessing the Value of Artificial Intelligence in Medical Imaging) [54], another translational-to-practice focused evaluation framework (Translational Evaluation of Health Care Artificial Intelligence framework) [55], another that is a publication appraising tool in radiology [56] to complement an existing one (Data Algorithm Training Output method) [57], another for setting specific “verification paradigms” of AI tools in clinical decision-making [58], one a function-specific (algorithm selection) [59], another that is ethics-focused [60], one forcedly matched to the sequential phases of experimental testing and clinical research for drugs and medical devices [61], while another is a reporting guideline developed through the Delphi method of 11 researchers from 3 institutions on 3 different continents, providing suggestions for every section of a journal paper from title to discussion on limitations [62]. There is another reporting guideline specific to radiomic research (Checklist for Evaluation of Radiomics Research) [63] for study planning, paper writing, and evaluation during the review process. There is one radiology AI software specification, classification, and performance evaluation framework till postdeployment, which is comprehensive but presented as fragmented parts and might complicate its application [64]. Another radiology framework (Radiology Artificial Intelligence Deployment and Assessment Rubric) emphasizes proper study designs in the assessment of its 7 preidentified function domains named as hierarchical levels of efficacy [106], and not along a clear development process of AI tools. A recent checklist, OPTICA (Organizational Perspective Checklist for Artificial Intelligence Solutions Adoption), was developed by a single institution primarily to assess completed AI solutions that are to be adopted in health care organizations [65]. It comprises 13 chapters of 3 to 12 items in each chapter to be completed by 5 main stakeholders who are assumed to be educated and able to mark off the checklist items competently and sequentially by a single identified stakeholder. This is different from ETEPAI, where full consideration is to be given by the project team at the beginning, and adhered to or referenced during the process till the end. Although OPTICA does not specifically address ethics, the 77 probing questions are indeed noteworthy to all stakeholders during adoption consideration. This is unlike ETEPAI, which mandates early and continuous adherence to ethical and procedural considerations throughout the entire AI project life cycle in general. TRIPOD-LLM (Transparent Reporting of a Multivariable Model for Individual Prognosis or Diagnosis-Large Language Models) [66] focuses on text-only large language model (LLM) projects and standardized reporting to enhance clarity, transparency, and accountability. While ETEPAI is embedded in project workflows, TRIPOD-LLM aids in appraising study quality and facilitating reproducibility with less emphasis on guiding processes from initiation to completion if used as a post hoc assessment tool. Another, which is named as Awesome AI Guidelines on GitHub, is a rich compilation of many policy documents, guidelines, and checklists related to AI, with the majority of them on ethics principles; only a few are dated beyond the year 2020 [107]. Although every evaluation framework has its respective strengths and should be used where appropriate, ETEPAI is believed to have all the pivotal aspects required of AI tools production, with an emphasis on methodology.

Addressing Implementation and Sustainability Challenges

Recent studies highlighted that current regulatory approvals for AI models often did not ensure fairness, with bias evaluation and mitigation typically occurring postdeployment [108]. Integrating AI models into clinical practice involves significant challenges, including information technology system integration, local fine-tuning, and addressing unintended consequences such as automation bias [108]. Robust human-machine collaboration and continuous monitoring are essential to improve performance and prevent failures caused by changes in software or equipment. Postdeployment maintenance and sustainability of AI models in clinical settings present significant ongoing challenges, including model drift, data drift, concept drift, harmful feedback loops, adversarial attacks, and potential biases that may emerge in real-world applications, retraining schedules, and managing overall AI tools maintenance [109]. Additionally, usability evaluation faces obstacles such as the lack of transparency and explainability in AI systems, often referred to as “black boxes” or “black box inside a black box” nature of commercial AI models [108], which can hinder user trust and satisfaction. To address these issues, it is essential to conduct usability evaluations according to recognized standards, assess user learning curves to inform training programs, evaluate stakeholder and patient acceptability to ensure alignment with user expectations, and redesign AI systems tailored for their intended clinical use, focusing on integration into workflows, reliability, and regulatory compliance [55].

Strategic Value and Generalizability

The ETEPAI framework is structured to proactively address these multifaceted challenges. By mandating that critical issues such as fairness, bias, and workflow integration be considered from the initial design stage, it shifts mitigation from a reactive, postdeployment afterthought to a foundational project requirement, including agentic AI uses [110]. Its dedicated postdeployment stage provides a clear roadmap for the continuous monitoring and usability evaluations essential for long-term sustainability and trust. This comprehensive, life cycle–based approach distinguishes ETEPAI from other valuable but more specialized guidelines. While built upon excellent existing frameworks such as TRIPOD-AI (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis - Artificial Intelligence, a reporting guideline), FUTURE-AI (Fairness, Universality, Traceability, Usability, Robustness, and Explainability - Artificial Intelligence, imaging-focused), and ALTAI (Assessment List for Trustworthy Artificial Intelligence, an ethics checklist), ETEPAI’s novelty lies in its integration of ethical, technical, and epidemiological considerations into a single, cohesive structure. Consequently, this end-to-end structure establishes its practicality as an actionable guide for multidisciplinary teams from project inception, rather than as a post hoc reporting or evaluation checklist.

Consequently, ETEPAI is designed to be model-agnostic. Whether the underlying technology is a traditional logistic regression, a convolutional neural network, or an LLM, the necessity for ethical oversight, rigorous epidemiological study design (such as external validation), and postdeployment monitoring remains constant. While the specific performance metrics may differ between model types, such as calibration plots for regression vs red-teaming for generative artificial intelligence (GenAI), the governance checkpoints provided by ETEPAI are universally applicable steps in the translational pathway.

Application Nuances

Notwithstanding the comprehensive and process-centric evaluation of ETEPAI, the degree of relevance and importance of ETEPAI’s items within the 4 stages of design, development, deployment, and postdeployment, and the 3 domains of ethics, technical, and epidemiological principles may differ in different AI projects of tools or systems. The different objectives, settings, and complexities of solutions of each AI tool would demand different attention on different items in ETEPAI. Nevertheless, every item requires consideration and must be justified if bypassed. When any of the items, stages, or domains raises concern, for proper understanding and evaluation, we are to refer to the original guideline or checklist (as given in the parentheses in the footnote to Table 1).

There are a few potential drawbacks of ETEPAI arising from the recommendations it is built on and the evolving techniques and technology of AI. Some of these include a lack of detail on the class imbalance mitigation strategy, approaches to address fairness, heterogeneity in estimates of model parameter values, and model performance. Additionally, further guidance will be needed on specific AI areas such as AI tools to influence social and health behaviors at scale in social media and different platforms, artificial general intelligence in health care, and other guides that were developed for specific purposes, as alluded to in this discussion, such as the OPTICA tool in assessing AI solutions in health care settings [65]. This is because ETEPAI is mainly for sign-posting the important considerations as necessary indicators and not for instructing the proper means of achieving them, which are dependent on the dataset’s quality and distribution. However, ETEPAI reminds developers and stakeholders to provide a clear explanation and reporting of all decisions taken during the whole process, from development to postdeployment. Guidance on advanced evaluation methods, such as handling Pareto frontiers for multiobjective optimization (balancing fairness and accuracy) [111,112], would require more elaborative and extensive explanation beyond the foundational structure ETEPAI aims to provide.

Limitations

While comprehensive, the ETEPAI may not fully address all aspects of certain types of AI models, especially newer or more complex models such as GenAI or agentic AI where multiple AI tools are developed and deployed to complement collaboratively [113], or in other sectors than health care such as education [114,115] and research enterprise [116,117], which are evolving and with robust evaluation frameworks that are yet available [118]. The POLARIS-GM (Partnership for Oversight, Leadership, and Accountability in Regulating Intelligent Systems–Generative Models in Medicine) [118] is an ongoing initiative focused on developing granular, scenario-based regulatory guidance for the unique challenges of GenAI and LLMs in medicine, using a multiphase, consensus-driven approach to generate new recommendations for postimplementation controls. In contrast, ETEPAI serves as a broad, comprehensive guide and evaluation tool for general AI tool production, synthesizing existing recommendations into practical pointers. We acknowledge that certain technical specifics of traditional ML do not map directly to GenAI, such as the standard calibration metrics, discrimination measures such as receiver operating characteristic area under curve, or the evaluation of feature stability, which are often replaced in GenAI by semantic similarity scores and human-preference alignment, and shift toward detecting toxicity and protecting against jailbreaking attempts, respectively. However, ETEPAI’s core principles remain applicable if adapted appropriately. These and the stages of considerations are still relevant to GenAI and agentic AI across their life cycle [119], from defining a safe operational domain in the design stage, and conducting adversarial “red-teaming” and error analysis to evaluate hallucination rates and consistency in the development stage, to mitigating automation bias through stakeholder education during the deployment stage and continuously monitoring for emergent harms in the postdeployment stage. While ETEPAI provides the overarching governance structure, developers must supplement it with specific, evolving guidelines for generative models, which incorporate initiatives such as the TRIPOD-LLM and the POLARIS-GM, to address their unique technical and ethical challenges.

When the referenced materials are used, the interpretation of criteria for items relies on expert ratings, which can introduce subjectivity and potential bias, leading to inconsistencies in the evaluation of AI tools across different studies and users. As this study is primarily on assimilating existing recommended materials to produce one overarching guidance and not on evaluating the quality of the referenced materials, there is neither analysis of their processes and products, nor were comparisons made among them. However, ETEPAI uses a consensus-driven approach by using the collective judgment of multiple experts, which balances individual biases and ensures a more reliable and balanced assessment. Additionally, it emphasizes transparency by documenting the criteria and rationale behind the ratings. This openness allows for thorough scrutiny and review, helping to identify and address potential biases or inconsistencies in the evaluation process.

All care was taken to screen through eligible guides and checklists at the time of publication of ETEPAI. There are many similarities among the included and excluded guides, but only those that are considered to be robust in development and focus on the process from design to postdeployment formed the basis of ETEPAI. However, the content of ETEPAI is noted to be inclusive of almost all important considerations in the excluded guides, as discussed. The similarity of ETEPAI to many guidelines and checklists is intended as per the method used to create it. As ETEPAI draws its validity from 30 robust, consensus-derived guidelines it synthesizes, this provides a strong evidence base for ETEPAI’s applicability across diverse AI models. Although the acceptability, applicability, and effectiveness of ETEPAI in real-world uses have not been tested, it is expected to be similar to those that were robustly tested. Although ETEPAI’s content validity is established through its systematic synthesis of existing, evidence-based, and widely accepted guidelines, making it conceptually sound upon proposal, it would require subsequent empirical work to confirm them. This may include methods such as (1) expert consensus: conducting a formal Delphi study with international experts in clinical AI, ethics, and epidemiology to refine and weigh the framework’s items; (2) pilot testing: implementing the ETEPAI framework within one or more real-world AI development projects to gather qualitative feedback on its practicality and utility; and (3) retrospective application: using ETEPAI as an evaluation tool for a cohort of published AI studies to assess its applicability and identify areas for improvement. While ETEPAI as a unified entity awaits prospective empirical testing, its component “data points,” the individual requirements for fairness, robustness, and reporting, are already established standards in the scientific community.

Despite its comprehensive nature, the ETEPAI may still be challenging for some stakeholders to fully grasp without sufficient training or motivation. This is evidenced by the tool’s expansive supplementary materials for full explanations, which may limit its immediate usability, as users need to refer to additional resources for a lucid understanding. Many pointers in the footnotes are cross-disciplinary and could be beyond any project team member. For example, the emphasis on diversity in the datasets and caution on using existing datasets that might limit the representativeness or comprehensiveness necessary for some AI tool development scenarios are to be appreciated by all stakeholders and upheld by methodologists or epidemiologists to ensure compliance throughout the whole production process. Although the framework highlights the need for rigorous external validation and adherence to reporting guidelines, these processes can be resource-intensive and may not be feasible for all projects. Furthermore, the rapidly evolving nature of AI technology, AI models updating methods [120], and health care practices may necessitate frequent updates to the ETEPAI framework to maintain its relevance and effectiveness. Public scrutiny and transparency are important. Publishing project protocols or making them available on public platforms can help ensure that the best practices recommended by ETEPAI are followed. It also makes it easier to explain and document any changes or deviations from the proposed design.

Conclusions

The ETEPAI is not just another guideline, but it combines existing robustly developed policies, guidelines, and checklists in one place that map out the ideal route, navigational aids, and risky areas throughout the whole AI tools development process. By focusing on its core function as a guide, this synthesized guideline encourages thoughtful application, ensuring its adaptability to diverse settings while avoiding misuse as a rigid standard, evaluation benchmark, or risk assessment framework. The guideline provides flexible principles and recommendations that must be adapted to specific contexts rather than being followed rigidly or universally applied. While it promotes quality, it does not include structured criteria or metrics for evaluating performance or outcomes. The guideline does not focus on identifying, quantifying, or mitigating risks but may complement such tools when integrated appropriately. Therefore, ETEPAI should be more useful as a training reference manual than a quality standard reference manual. This is because to appreciate and to put it to immediate practice requires previous training on AI concepts and technologies. It is possible to use ETEPAI as a template for training purposes to impart AI-related clinical competencies in health care professionals [121]. It is both comprehensive and complete, containing almost all critical considerations, as well as briefly and crisply presenting those considerations in a table of probing questions and another table of critical pointers in 3 domains of ethics, technical, and epidemiological principles when considering an AI tool.

Applying ETEPAI ensures comprehensive guidance and encourages a multidisciplinary approach to AI tool development, from design through postdeployment. This framework promotes adherence to ethical standards, robust technical execution, and rigorous epidemiological research methodologies. By integrating these considerations, AI tools can achieve their intended impact in health care settings, offering reliable, effective, and ethically sound solutions. To ensure effective application of ETEPAI and meticulous attention to the aspects throughout the AI development life cycle, it is essential for users to understand each and every included guideline, checklist, assessment, and framework, and be familiar with the ethical and regulatory concepts. We highly advocate its use to ensure that developers and authors appropriately design and comprehensively report AI work leading to high-quality, practical AI applications that align with clinical needs and ethical imperatives. ETEPAI’s product-centric approach may effectively align with EU trustworthiness benchmarks on safety, robustness, and ethics from design to deployment [13]. This helps the final AI products to be compliant with the EU’s requirements for the protections of health, safety, fundamental rights, and comprehensive risk management that considers impacts on both products and systems [75]. Using ETEPAI should be easy and sufficient for experienced professionals and serve as clear signposts for less prepared stakeholders. As the field of AI continues to evolve with new technologies and algorithms, we anticipate ongoing updates to ETEPAI as new guidelines and reporting standards emerge to ensure high-quality scientific research and the ethical application of AI in medicine.