Artificial intelligence (AI) has emerged as a transformative technology with vast potential across various sectors, including health care [1,2]. In this field, AI-based medical devices have garnered significant attention due to their ability to revolutionize diagnosis, treatment, and patient monitoring [3]. These devices use advanced algorithms and machine learning techniques to analyze complex medical data sets, thereby providing valuable insights and support to health care professionals in their decision-making. To meet the ever-increasing demand for integration of AI-based medical devices into clinical practice, their efficient evaluation through adapted health technology assessment (HTA) is now crucial [4]. Several frameworks have been published showing an adaptation of items for reporting clinical trials related to AI-based medical devices (Table 1). However, there is still a lack of a much needed adaptation of the standard HTA framework to better suit the assessment of AI-based health care technologies.

An HTA involves a systematic evaluation of a medical technology’s clinical, economic, ethical, and social implications to determine its overall value and impact on health care delivery [9]. The European Network for Health Technology Assessment has designed an HTA Core Model that provides a methodological framework for production and sharing of HTA information [5]. It evaluates the following nine domains: (1) health problem and current use of technology, (2) description and technical characteristics of the technology, (3) safety, (4) clinical effectiveness, (5) costs and economic evaluation, (6) ethical analysis, (7) organizational aspects, (8) patients and social aspects, and (9) legal aspects (Multimedia Appendix 1).

A full understanding of the capabilities, limitations, and specificities of AI-based medical devices is paramount to complete these HTA domain assessments and, thereby, inform evidence-based decision-making and allow for policy development and the responsible integration of these technologies into health care systems [10,11]. Much uncertainty remains with regard to the reliability of AI-based medical devices, data issues, and regulatory processes, resulting in multiple challenges faced by HTA agencies assessing new technologies and delivering their approval [4]. Nevertheless, AI-based medical devices require strict regulations and specific legislations [12-14]. Over the last few decades, the evaluation of AI-based medical devices through an HTA process has received growing interest, as shown by the increase from 1 published article in 1990 to 94 in December 2023, with 484 articles in total during the period 1990-2023 (Table 2).

The objective of this review was to critically assess the comprehensive suitability of the current HTA process for AI-based medical devices. By evaluating the performance and capabilities of AI-based medical devices across multiple dimensions, this review aimed to inform health care professionals, policy makers, and researchers about the challenges and opportunities associated with these technologies. Ultimately, this review sought to facilitate evidence-based decision-making; promote responsible implementation; and maximize the potential benefits of AI-based medical devices in improving health care quality, accessibility, and outcomes.

To this end, we analyzed the suitability of each HTA domain for the assessment of AI-based medical devices and proposed an adapted HTA framework.

A scoping literature review was conducted following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) methodology [15,16]. All articles related to HTA methods for AI-based medical devices were selected. Data extraction focused on assessment criteria, methodological evaluations, and results.

The search terms and strategy outlined in Textbox 1 were used for the scoping review.

We searched multiple databases: PubMed, Embase, and the Cochrane Library. Additional articles were retrieved manually from the gray literature and from HTA agency websites (Bundesinstitut für Arzneimittel und Medizinprodukte [Federal Institute for Drugs and Medical Devices in Germany], Haute Autorité de Santé [French Health Authority], National Health Service, National Institute for Health and Care Excellence, National Institute of Public Health, International Network of Agencies for Health Technology Assessment, and the European Network for Health Technology Assessment).

The studies were selected by 2 reviewers (LF and AV). After removing duplicates, both reviewers independently screened the abstracts to select eligible articles and then analyzed full-text reports for eligibility. A third party (NM) resolved the possible discrepancies highlighted during the selection process should a consensus not be reached. An extraction database was used to list the selected articles meeting the inclusion criteria to ensure that all eligible articles were included.

The inclusion and exclusion criteria were related to (1) language (this review was limited to the English and French languages), (2) article type (reviews and primary research were included, whereas other article types were excluded, such as abstracts, commentaries, editorials, letters to the editor, case reports, case series, animal studies, phase-1 and phase-2 studies, pilot studies, duplicate studies, irrelevant studies, studies with the wrong aim, and studies available in abstract form only), (3) type of technology (only AI-based medical devices were eligible), and (4) type of evaluation (HTA articles were included).

In total, 2 analysts (LF and AV) independently extracted data items from the selected articles. A third party was involved to resolve any discrepancies highlighted during the selection process. The following data were extracted from each article: (1) general characteristics of the studies (authors, journal, and publication date), (2) study objective, and (3) HTA assessment domain for AI-based medical devices related to the article.

Neither the methodological quality nor the risk of bias of the included articles were assessed, consistent with scoping review guidance [15].

Concerning each HTA domain, we added a case study of the HTA of several AI-based medical devices in diabetes to illustrate our recommendations.

The literature search resulted in 775 citations summarized in the PRISMA flowchart (Figure 1). After removal of 32.6% (253/775) of duplicates, we carried out an initial screening of the remaining 522 publications, resulting in 129 (24.7%) potentially relevant full-text papers. After further screening, we assessed 77 reports for eligibility, based on which we excluded 11 (14%) for not referring to AI-based medical devices and 2 (3%) for not being HTAs. We then included a further 14 records from 59 potential citations available from HTA agency websites and gray literature.

This gave a total of 78 included articles covering one or multiple HTA domains, the distribution of which is provided at the bottom of Figure 1.

We then summarized the data collection for each selected article in Multimedia Appendix 2 [1,2,4,9,10,17-89].

As domain 1, health problem and current use of technology, is suitable for any type of medical technology and systematically addressed these technologies, we focused on the suitability of the HTA domains (from 1 to 9).

Taking into account the previous considerations, some recommendations could be proposed to adapt and personalize these standard HTA domains to AI-based medical devices’ specificities. Therefore, we suggest 4 main recommendations by domain in Table 3.

The adaptation of the HTA process through a methodological framework for AI-based medical devices enhances the comparability of results across different evaluations and jurisdictions [90-92]. It promotes consistency in the assessment methodologies, reporting formats, and presentation of findings, enabling decision makers to make informed choices based on reliable and comparable evidence. This standardization contributes to the overall credibility and acceptance of HTA outcomes related to AI-based medical devices [93]. In addition, a methodological framework would address the need for appropriate expertise and skills to conduct the HTA of AI-based medical devices [94-99]. It would outline the qualifications and competencies required for the individuals involved in the assessment, including knowledge of AI technologies [46]. By defining the necessary expertise, the framework supports the development of a skilled workforce capable of conducting robust and reliable HTAs of AI-based medical devices (Figure 2).

Evaluating the trade-offs and weighing different features of AI-based medical devices is indeed a complex and important task, especially in domains such as health care where the impact on human lives is significant. The acceptability of AI-based medical devices should be assessed on a case-by-case basis considering various factors, including, for instance, performance, accuracy, cost, explainability, and the specific context in which they are being used.

Trade-offs between accuracy and cost are common in AI. It may be acceptable for AI to increase follow-up care costs if it significantly improves accuracy and patient outcomes. For example, if an AI system can detect diseases at an earlier stage, it might lead to more effective treatment and, ultimately, lower overall health care costs in the long run.

The balance between explainability and performance is a critical consideration. While explainable algorithms are preferred for safety and transparency reasons, there may be cases in which a highly complex, unexplainable algorithm outperforms explainable ones. In such cases, the trade-off between transparency and performance should be carefully evaluated based on the specific use case and potential risks.

The minimum performance for adding this technology to the tools available to human clinicians depends on the specific task and the level of trust that patients, health care providers, and regulators have in the technology. Some key factors to consider include the following: (1) The complexity of the task—AI-based medical devices should excel in tasks that are well defined and data driven, but they may not replace human clinicians in tasks requiring complex decision-making, empathy, or ethical considerations. (2) Safety and reliability—AI-based medical devices should demonstrate a high level of safety and reliability, ideally surpassing the performance of human clinicians in terms of avoiding errors. (3) Ethical considerations—AI-based medical devices should adhere to ethical standards, including patient privacy, informed consent, and unbiased decision-making, which are often considered even more important than performance metrics. (4) Regulatory approval—regulatory bodies often establish performance thresholds for AI-based medical devices. Compliance with these thresholds is essential for market acceptance.

In general, AI should aim to complement and enhance the capabilities of human clinicians rather than completely replacing them. The specific threshold for acceptable performance will vary across applications and contexts, and it should be determined through a combination of rigorous testing, peer-reviewed studies, and input from health care professionals and patients.

It is important to note that ethical considerations, patient safety, and the potential for bias should always be at the forefront of these discussions, and AI-based medical devices should not be adopted solely for the sake of automation or cost reduction if they compromise these critical aspects of health care.

Concerning the HTA domain 3 on the safety of an AI-based medical device, the potential risk of patient injury of an insulin delivery AI-based medical device system should be taken into account in the risk management process of algorithm development [96]. In the case of an evolutive deep learning–based medical device without continuous ongoing safety assessment, a wrong dosage administration due to an AI error, for instance, could provoke a serious adverse event such as an acid ketosis coma for a patient with diabetes.

A risk management plan should be available for users and regularly updated with the impact of the AI changes on the safety plan.

In relation to the HTA domain 4, which focuses on the effectiveness of an AI-based medical device, the FDA in the United States has proposed a regulatory framework for modifications to evolutive AI- and machine learning–based software as a medical device with modification guidance focused on the risk to users or patients resulting from the AI changes [100]. They have asked for an Algorithm Change Protocol with specific methods in place to achieve and control the risks of the anticipated types of modifications related to performance, use, or inputs. A continuous clinical effectiveness assessment could be an interesting approach for an AI-based medical device detection system for diabetes foot ulcer for patients needing adaptative treatment modifications to prevent ulcer development [101].

Concerning the HTA domain 5, to evaluate a diabetic retinopathy screening AI-based medical device, clinical effectiveness is not sufficient. The cost-effectiveness of different types of AI diabetic retinopathy screening should be compared with no screening and ophthalmologist screening. A recent study demonstrated that AI-based screening was the most cost-effective, not only saving costs but also improving the quality of life of patients with diabetes [44]. In this case, the long-term assessment of the economic impact of AI introduction in diabetic retinopathy screening highlighted the added value of this technology.

The HTA domain 7 on organizational impact has to assess how the AI-based medical device can be effectively integrated into the health care pathway and prevent wasteful spending. More thorough attention must be paid to the following aspects: (1) evaluating needs and determining the added value of the implementation of the AI-based medical device; (2) assessing workplace preparedness, including stakeholder acceptance of the introduction of the AI-based medical device and involvement; and (3) analyzing the alignment between AI technology and organizational structure [64]. Decision makers and technology advocates need to address the complexities of AI more comprehensively and understand the systemic challenges that its adoption poses to health care organizations and systems. As an example, consider an AI tool used for diagnosing diabetic retinopathy in a primary care setting, such as by a family physician or nurse [102]. In theory, this could lead to shorter waiting times for patients. However, if the health care organization faces challenges such as a shortage of specialized staff (eg, ophthalmologists), insufficient organization of care pathways, and lack of specialized facilities for proper management and follow-up after diagnosis, the introduction of AI might adversely affect the quality of care and patient experience. In such a scenario, the AI application might merely transfer the delay from primary to secondary care, failing to address the fundamental issue.

Finally, concerning the last domains (6, 8, and 9) about ethical, patient, social, and legal aspects, AI-based continuous glucose monitoring and insulin pumps should also be assessed on ethical and legal aspects [103]. While citizen juries have generally shown support for AI in research and treatment, concerns remain. The risks of data theft and privacy breaches necessitate careful consideration of ethical and legal issues for patients. Although AI can aid in decision-making, it cannot wholly substitute a physician’s expertise. Effective regulations and systems designed to ensure safety, reduce bias, and enhance transparency are essential.

By implementing these recommendations, stakeholders can foster the responsible integration, regulation, and evaluation of AI-based medical devices. These measures can enhance the evidence base, address ethical concerns, and maximize the potential benefits of these AI technologies in improving health care outcomes while protecting patient safety, privacy, and equity.

When it comes to implementing specialized HTAs for AI-based medical devices, several practical challenges and potential barriers could be encountered and should be taken into account: First, AI technological evolution timeline—AI technologies evolve at a much faster pace compared to traditional medical devices, making it challenging for HTA frameworks to keep up with the latest developments and assess their long-term impact effectively. AI-based medical devices have a short product lifetime, between 12 and 18 months, in contrast to drug products. This shorter life cycle highlights the need for evolutive and fast-track HTA processes for AI-based medical devices. Second, data requirements and quality—AI systems rely heavily on large data sets for training and validation. Ensuring the availability of high-quality, representative data is a significant challenge. There is also the issue of data privacy and security, which must be addressed. The availability and quality of data and evidence required for conducting HTAs on AI-based medical devices present a complex and evolving landscape. Assessing the feasibility and challenges associated with gathering such data is crucial for robust evaluations. First, the availability of data can vary significantly depending on the AI-based medical device. While some devices may have access to vast amounts of high-quality real-world patient data, others might face limitations due to the novelty of the technology or issues related to data privacy. Second, the quality of data is paramount as inaccurate or biased data can lead to flawed assessments. Ensuring data accuracy, representativeness, and relevance is a constant challenge in AI-based medical device evaluations [104]. The rapid pace of AI development can result in limited long-term data, making it difficult to assess the device’s sustained performance and safety [105]. Balancing the need for robust evidence with the dynamic nature of AI technologies is a significant challenge that HTA organizations must address to provide valuable insights for informed decision-making in health care. Third, complexity and transparency—the complex algorithms used in AI-based medical devices can be difficult to understand and assess, leading to issues with transparency and explainability [17]. This complexity can pose a challenge for regulators and assessors in HTA processes. Fourth, clinical validation and safety requirements—generating robust clinical evidence to demonstrate the safety, efficacy, and effectiveness of AI-based medical devices can be challenging. This includes proving that these devices perform consistently across diverse patient populations. Moreover, to account for a life cycle that would include updates that may improve performance, one solution to investigate could be the proposition of the FDA in the United States in 2019 to experiment with a “dynamic certification” [100]. It allows for the re-evaluation of the AI-based medical device in case of a substantial modification of the indication of the medical device or the way to deliver the diagnosis, for instance. Fifth, regulatory and ethical considerations—adapting existing regulatory frameworks to accommodate AI-based medical devices, addressing ethical concerns such as bias, and ensuring equitable access are critical challenges [106]. Sixth, economic evaluation—determining the cost-effectiveness of AI-based medical devices, especially when benefits might be indirect or long term, poses a unique challenge for HTA [93]. Seventh, stakeholder engagement and trust—building trust among health care providers, patients, and policy makers regarding the reliability, trustworthiness, and usefulness of AI-based medical devices is crucial but challenging [107]. Eighth, integration into health care systems and interoperability—the integration of AI-based medical devices into existing health care workflows and systems can be complex and resource intensive [18]. Ninth, global and local applicability—ensuring that AI-based medical devices are effective and appropriate for use in different global and local contexts, considering varying health care systems and population needs, is another significant barrier.

Practically, one suggestion to implement such frameworks could be to implement these recommendations in the future European clinical joint assessment guidelines. As they are being currently discussed in a European project on a common HTA process for connected medical devices that includes AI-based medical devices, it could be the opportunity to tackle these challenges at the European level [108].

Addressing these challenges requires a concerted effort from regulators, health care providers, technology developers, and other stakeholders in the health care ecosystem.

AI-based medical devices have the potential to transform health care delivery, but the suitability of the current comprehensive HTA requires careful adaptation of the evaluation across the 9 dimensions. While these AI devices show promise in improving accuracy, safety, and efficiency, there is a need for robust clinical validation, integration into workflows, economic evaluation, and addressing of ethical and legal implications. A comprehensive adapted HTA framework for AI-based medical devices can provide valuable insights into their effectiveness, cost-effectiveness, and societal impact, guiding their responsible implementation and maximizing their benefits for patients and health care systems.