Introduction
Recent advancements in artificial intelligence (AI) have demonstrated successful results in various clinical practices, and there is increasing anticipation of AI-based technologies addressing the global healthcare crisis. This crisis arises from a shortage of healthcare professionals, aging populations (Reference Ma, Yang and Wong1), and limited financial resources (Reference Ciecierski-Holmes, Singh, Axt, Brenner and Barteit2).
AI encompasses a wide range of applications and technologies. As defined by the Organisation for Economic Co-operation and Development (OECD), AI is “a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that [can] influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment” (3;4;Reference Estevez Almenzar, Fernandez Llorca, Gomez Gutierrez and Martinez Plumed5).
In the context of health, AI-based technologies serve as an umbrella term, including machine learning algorithms and other cognitive technologies that utilize medical data to automate specific tasks. These applications aim to cover the entire patient journey, supporting clinicians in diagnosis, therapeutic decision making, and predictions (6).
Globally, the healthcare market size of AI is estimated to be USD 15.4 billion in 2022, with an expected compound annual growth rate of 37.5 percent from 2023 (7). In Europe, as well as in other parts of the world, health has been recognized as a key application for AI. Nevertheless, a significant need for AI regulatory frameworks and a code of practice to address healthcare-specific risks and requirements has been highlighted. This recognition is evident in various documents, including the European Strategy on AI from 2018 (8), the Guidelines for Trustworthy AI in 2019 by the High-Level Expert Group on AI (9), and the recently proposed legal framework on AI, which is the first of its kind, introduced in April 2021 (10). Notably, the European Parliament has recently released the first regulation on AI (11).
The implementation of new medical technologies and clinical pathways in health care is firmly grounded in research and an evidence-based scientific approach. It is a standard practice for clinicians to rely on health technology assessment (HTA) methods to aid in the decision-making process regarding the adoption of new technologies (Reference O’Rourke, Oortwijn and Schuller12). The purpose of HTA is to support and inform policy decision making based on a systematic and evidence-based approach. In Europe, the primary reference framework for HTA is the European Network for HTA (EUnetHTA) Core Model (13), which guides assessors through a comprehensive evaluation of technologies across nine different domains. This model has been developed and adopted for medical devices and pharmaceutical products, enjoying broad recognition of its value among relevant stakeholders including industry (Reference Gyldmark, Lampe and Ruof14;Reference Alami, Lehoux and Auclair15). However, AI-based health technologies (AIHTs) challenge the applicability of traditional HTA methods due to innovative technologies evolving at a pace faster than the methods used for conducting HTA. Key challenges posed by AIHTs from the HTA perspective include the following:
-
• Nature of AI (Reference Bélisle-Pipon, Couture and Roy16), given that these types of technologies may (as is the case with adaptive algorithms) or may not continue to evolve.
-
• Lack of transparency and replicability (Reference Vollmer, Mateen and Bohner17).
-
• Ethical and legal implications are widely debated both in a general context (Reference Naik, Hameed and Shetty18) and concerning specific clinical applications, such as breast cancer (Reference Carter, Rogers and Win19), among others.
Furthermore, the limitations or inability of most AI technologies to “explain” their decision-making process underscores the importance of updating traditional HTA methods. This update should include new aspects such as trustworthiness (20;Reference Zhang and Zhang21), transparency, interpretability (Reference Reyes, Meier and Pereira22), and explainability (Reference Combi, Amico and Bellazzi23;Reference Amann, Blasimme and Vayena24) within the HTA framework. These additions are crucial to provide decision makers with the proper support when considering the adoption of AI (Reference Farah, Davaze-Schneider and Martin25).
Despite attempts to align HTA methods with AI adaptation, such as the Model for ASsessing the value of AI in Medical Imaging (MAS-AI) (Reference Fasterholdt, Naghavi-Behzad and Rasmussen26), Digi-HTA in Finland (Reference Haverinen, Keränen and Falkenbach27;Reference Haverinen, Turpeinen, Falkenbach and Reponen28), AQuAS Framework for digital health (Reference Segur-Ferrer, Moltó-Puigmartí, Pastells-Peiró and Vivanco-Hidalgo29), and the evidence standards framework adopted by NICE for digital health technologies (30), no joint agreement among experts exists on how to assess the value and effect of AI-based technologies.
Therefore, the objective of this study is to examine the perception among European healthcare decision makers, assessors, and experts regarding the usability of the EUnetHTA Core Model for assessing AI-based technologies. Additionally, the study aims to explore their perception of including new assessment topics identified in the literature as important for the assessment of AI-based technologies.
This study is being conducted as part of the European Union (EU) project “The AI-Mind” (31), supported by the European Research and Innovation Action Plan (No. 964220). The main objective of the EU project is to develop AI-based diagnostic tools for early screening and risk assessment to predict the onset of dementia (32). Subsequently, the study aims to evaluate the usability of the developed tool (Reference Haraldsen Ira, Hatlestad-Hall and Marra33;34). In the absence of HTA frameworks that support the evaluation of AI-based technologies, it has been decided, as a first step, to initiate an early dialog among stakeholders. This dialog will involve patients, developers, industry representatives, clinicians, and HTA experts, with the purpose of setting priorities and identifying the evidence required to inform decision-making processes.
Methods
Our analysis was conducted in three steps. First, a list of potential HTA topics relevant to AI was identified based on a scoping review. Subsequently, a Delphi survey was conducted among qualified experts, and our approach adhered to the guidelines for the Delphi survey (Reference Hasson, Keeney and McKenna35;Reference Hsu and Sandford36).
Step 1: Identification of HTA Issues Relevant for AI
An initial catalog of HTA domains and topics was created based on the EUnetHTA Core Model (version 3.0) (13). Following this, a rapid review was conducted to identify additional relevant topics for inclusion. This involved searching both scientific and gray literature across platforms such as PubMed, Cochrane Library, Google Scholar, and websites of major HTA agencies. Additionally, we considered abstracts from HTAi annual meetings, the ISPOR Presentations Database, and INAHTA member resources.
The syntax used for the literature review in PubMed is provided in Supplementary File 1a, and the websites of the HTA agencies included in our analysis are listed in Supplementary File 1b. Furthermore, we identified reports covering AI topics in health from international institutions such as the World Health Organization (WHO), the European Commission, and the OECD.
Step 2: Delphi Survey
Topics identified in Step 1 were used in the development of the Delphi survey. A modified version of the Expert Delphi technique was selected for consensus building, and it involved the following four steps: (i) Development of an online survey. (ii) Recruitment and consenting of participants to the Delphi panel. (iii) Two rounds of consultation on the proposed topics in the survey. (iv) A webinar for the expert panel.
Design
The survey covered the nine domains and associated topics presented in the EUnetHTA Core Model and additional topics identified through the rapid review. In total, the survey comprised sixty-five multiple-choice questions. Each topic was briefly described, and after each domain, a free-text question was included for the panelists to provide comments. The full list of domains and related topics is reported in Supplementary File 2.
Information regarding occupation, expertise, knowledge about HTA and AI, and geographical location was also collected from each panelist.
For the consensus process of the Delphi survey, panelists used a 9-point Likert scale to rate each statement. A score of 9 indicated the highest level of agreement for inclusion, whereas a score of 1 suggested that the topic should not be included in an HTA of AI-supported technologies. The survey specified that a score from 1 to 3 should be interpreted as “should not be included in an HTA on AI,” a score from 4 to 6 as “important but not critical to include in HTA of AI,” and a score higher than 7 as “critical to include in HTA of AI.”
The electronic survey was prepared using the user-friendly Alchemer online interface (www.alchemer.com).
We used 70 percent as a cutoff for high agreement among experts for each topic. This cutoff was applied not at a single point on the Likert scale but for each of the three categories described above (critical to include, important not critical, and not important).
The choice of the cutoff was based on guidelines (Reference Nasa, Jain and Juneja37) and works from the WHO (38;Reference Cascella, Miceli, Cutugno and Morabito39;Reference Tejwani, Chang and Tran40). For example, if ≥70 percent of the responses fell in the range of seven to nine, the topic was considered critical and to be included in the assessment of AIHTs.
Participants
Our goal was to assemble a multidisciplinary expert panel.
Potential participants for the Delphi panel were identified based on their publications, CVs, area, and level of expertise. In terms of geographical representativeness, the primary focus was on Europe-based experts. We identified nine relevant categories of experts: (i) clinician/researcher; (ii) HTA; (iii) technical experts (e.g., data programmer/engineer and cybersecurity); (iv) ethicist/bioethicist; (v) patients/advocates; (vi) health economy; (vii) health policy; (viii) legal aspects; (ix) user experience. To ensure representation from all categories in the Delphi process, we aimed to have a minimum of five representatives from each main expert group (groups 1–2–3-6) participating in the survey. Anticipating a response rate in the range of 30–40 percent, a total of 87 experts were invited.
Data Collection
Experts were invited to participate in the Delphi survey through email, which outlined their expected involvement and rights as participants, along with a link to the online survey. We conducted a two-round online Delphi survey. The second round duplicated the questions from the first round, but participants could view ratings (percentage of respondents according to importance score) from the initial round. This allowed them to adjust, confirm, or reconsider their answers. Finally, the expert panel was invited to a webinar to discuss the results of the survey and provide additional feedback on the survey, AI technologies, and HTA models for AI.
Step 3: Statistical Methods
The analysis was conducted using Microsoft Excel, with key indicators focusing on the proportion/percentage of respondents based on the importance score and category for inclusion/exclusion from the HTA–AI framework.
In this paper, we present results obtained at the conclusion of the second round of the Delphi survey, along with findings gathered from the discussions held during the virtual workshop.
Results
The Delphi survey spanned from April 2022 to January 2023, with the final webinar held in May 2023. Of the 87 experts invited, 46 responded to both rounds of the Delphi survey (Supplementary File 1c), resulting in a response rate of 47.4 percent, exceeding the anticipated value.
The majority of respondents (n = 43, 93.5 percent) were from Europe, representing fourteen different European countries. Italy (n = 14, 30.4 percent) and Norway (n = 6, 13 percent) demonstrated particularly high response rates (see Figure 1). Additionally, three panelists were based in other continents, specifically Canada, Tunisia, and Australia.
Figure 1. European Union countries represented in the panel.
As outlined in Table 1, 12 panelists (26.1 percent) were clinicians and 7 (15.2 percent) were HTA experts. The majority of panelists reported having prior knowledge of HTA (n = 37, 80.4 percent) and familiarity with the EUnetHTA Core Model (n = 28, 60.9 percent). A smaller number of panelists (n = 5, 10.9 percent) reported having competencies in all HTA domains. Twenty-one (45.7 percent) possessed expertise in the assessment of clinical effectiveness, and 20 (43.5 percent) had experience in costs and economic evaluation. Regarding the ethical, legal, and social implications domains, only three experts had experience in assessing legal aspects, whereas 16 (34.8 percent) had experience in the other two domains, as well as in evaluating the organizational impact of health technologies.
Table 1. Panel composition and expertise
Approximately 65.2 percent (n = 30) of panelists reported at least a moderate level of familiarity with AI. Additionally, sixteen responders had practical experience in assessing or implementing AI-supported health technologies/solutions (Table 1).
Delphi Panel Results
Overall, the Delphi panel expressed agreement on the importance of including 73.8 percent of the original topics suggested in the Core Model (refer to Figure 2) and 70 percent of the additional topics (refer to Figure 3) identified in the literature for the assessment of AI technology (deemed critical to include in HTA of AI). Further details for each topic are provided in Supplementary File 3.
Figure 2. Summary of results for each traditional domain according to the EUnetHTA Core Model.
Figure 3. Summary of results for additional topics.
Concerning the first domain, “health problem and current use of the technology,” our experts unanimously agreed on the inclusion of all topics in the HTA framework for AI. A similar consensus was reached for the four topics falling under the domain labeled “description and technical characteristics of technology,” although agreement on the Regulatory Status was close to the cutoff value, with 71.74 percent of panelists assigning a score of ≥7.
In the safety domain, our experts concurred on the inclusion of only two topics. Divergent perceptions were noted regarding the relevance of occupational and environmental safety. For clinical effectiveness, only one topic (“change-in-management”) fell into the exclusion area based on our criteria for assessing agreement.
Despite varying levels of agreement, all topics related to cost and economic evaluation emerged as crucial for inclusion in the framework. The lowest level of agreement was recorded for the characterization of uncertainty (70.45 percent) and heterogeneity (71.11 percent).
The importance of ethical aspects was evident from the collected responses, with no topics excluded. For organizational aspects, diverse scores were assigned, with agreement reached only for the cruciality of the topic “process-related costs” (75.56 percent). A similar pattern was observed for patient and social aspects, particularly for the topic of patients’ perspectives.
In the legal aspects, at least 70 percent of panelists considered 5 out of 7 topics crucial.
Figure 3 clearly shows how the proposed twenty additional topics captured crucial aspects of AI in the majority of cases (n = 14).
Upon reviewing the overall percentage of scores higher than seven, our panel identified the ten most crucial topics to incorporate into an HTA framework for AI-based technology. These topics are as follows: accuracy of AI model (97.78 percent), patient safety (95.65 percent), evaluation of benefit–harm balance from an ethical perspective (95.56 percent), function (95 percent), target condition (93.48 percent), technology features (93.48 percent), risk management (93.48 percent), evaluation of benefit–harm balance in the clinical effectiveness domain (93.48 percent), data bias (91.30 percent), and measurement and estimation of outcomes (91.11 percent).
By combining these findings with the proportion of topics to be included in the HTA framework for each domain, a ranking of domains was established, as depicted in Figure 4. Ethical analysis emerged as the most relevant domain, occupying the pinnacle of the pyramid, whereas organizational aspects appeared at the base as the least critical domain.
Figure 4. Hierarchy of health technology assessment domains as perceived by panel of experts.
Final Webinar
The results were communicated to all experts via email, and seventeen experts participated in the final webinar. During the 60-minute online meeting, feedback was collected regarding the results and the adopted methodology. First, the discussion underscored unanimous agreement among experts on the necessity for an AI-adapted HTA framework. It was emphasized that this framework should not be limited to the EUnetHTA Core Model but should instead serve as a flexible starting point. The recognition of the value of including additional topics was unanimous among panelists. Data-related topics—such as bias, acquisition, and application—are widely regarded as critica l in AI development, along with broader issues like human agency, oversight, and explainability.
The experts stressed the importance of the AI-adapted framework being flexible enough to accommodate the heterogeneity of AI technologies. They pointed out that AI encompasses a diverse array of technologies with applications ranging from primary/secondary prevention to diagnosis and patient management. It involves various end users such as clinicians, healthcare professionals, and patients. Thus, the relevance of certain topics may vary across different AI health technologies. The panel recognized the need to adapt the Core Model to specific types of technology and decision-making contexts, acknowledging the validity of the process that guided the survey’s definition.
However, the panel identified potential biases and limitations. Varying levels of expertise in HTA and AI among panel members could have influenced the results, impacting the interpretation of proposed topics. The lack of a detailed definition for each topic in the survey, particularly for Core Model topics, might have presented a “cultural” barrier for experts not aligned with the EUnetHTA evolution and tools. Conversely, experts in traditional HTA assessments might not have been familiar with new AI-related topics, such as learning and training of models. Additionally, the development of a specific HTA terminology for AI is still in progress, leading to potential differences in the interpretation of terms and topics like interpretability or trustworthiness.
Furthermore, some of the proposed topics are relatively new in HTA, and the inclusion of environmental impact assessment is particularly significant and challenging. Nevertheless, the environmental impact of AI has been already proved in terms of energy cost and related carbon emission. For instance, the carbon footprint of training a single big language model was estimated equal to around 300,000 kg of carbon dioxide emissions (Reference Dhar41). Lack of experience and common methodologies may have contributed to the lack of agreement on this topic.
Finally, the results revealed some overlap among domains and topics in the EUnetHTA, as seen with the benefit–harm balance. However, it was acknowledged that topic overlapping is a well-known challenge of the Core Model and can be resolved during the adaptation phase.
Discussion
Our analysis was based on a strong assumption – the widely recognized need for a dedicated framework for HTA in AI. We explored the specific levels of agreement or disagreement regarding the topics to be covered in an HTA–AI framework.
We presented an extensive list of potential relevant topics to a panel of experts in HTA and AI, including those listed in the EUnetHTA Core Model and additional ones. Through the Delphi survey, we were able to assess the level of agreement for each of these sixty-five topics, categorizing them into three groups: critical to include, important but not critical, and not important. By applying our criteria (≥70 percent of responses in the same group), the experts reached a consensus that forty-eight (73.8 percent) of the topics are critical and should be an integral part of an HTA framework for AI-based technologies.
Our findings highlighted the inadequacy of the Core Model in capturing all relevant topics for AI-based technologies. Experts unanimously supported the inclusion of 14 out of the 20 additional proposed topics in the HTA–AI framework. Interestingly, two of these topics – accuracy of the AI model and bias in data – were among the top ten most critical.
Furthermore, our analysis revealed the importance of ethical aspects in AI, placing them on the same level as, or even higher than, considerations of clinical effectiveness (see Figure 4). These results are consistent with other studies that identify new ethical, legal, and social challenges for the assessment of AI, with a focus on issues such as trust among clinicians and patients, as well as autonomy (Reference Bélisle-Pipon, Couture and Roy16).
Strengths
The study outlined in the paper distinguishes itself from other HTA frameworks proposed in the literature, ensuring that it is not a duplication. For example, while MAS–AI (Reference Fasterholdt, Naghavi-Behzad and Rasmussen26) shares a similar methodology – employing literature reviews and expert involvement – it is specifically tailored to medical imaging within the Danish context. In contrast, our perspective is more expansive, transcending limitations related to imaging or specific clinical indications. As highlighted in the paper, the Delphi panel, conducted as part of the AI-Mind EU project, intentionally avoided confining the research to dementia or Alzheimer’s disease. Despite a significant number of Italian experts (n = 14) in the final panel, our outreach spanned 15 EU and 3 non-EU countries (refer to Figure 1). Furthermore, our investigation into both the AI and HTA backgrounds of experts revealed that 65 percent of them were at least moderately familiar with AI (Table 1), and 80 percent had prior knowledge of HTA. These insights served to identify biases and limitations, guiding us toward potential areas of improvement. Notably, although there is a certain degree of overlap in results when comparing the MAS–AI domains (Reference Fasterholdt, Naghavi-Behzad and Rasmussen26) with our Figures 2 and 3, it underscores how some peculiarities of AI are universal across clinical applications. Looking forward, as the AI-Mind proposes an HTA–AI framework, experiences like MAS–AI will be invaluable in the adaptation process. During the final webinar, our experts affirmed the necessity for this adaptation, considering not only clinical applications but also expected end users and the technology readiness level (TRL) of the AI solution.
From the HTA perspective, the TRL (42) is significant in the assessment process (Reference Fasterholdt, Lee, Kidholm, Yderstræde and Pedersen43). The AI-Mind platform is set to introduce two new AI-based tools: the AI-Mind Connector, which identifies dysfunctional brain networks through high-density electroencephalographic recordings, and the AI-Mind Predictor, which assesses dementia risk using data from the Connector. These data include advanced digital cognitive tests, genetic and protein biomarkers, as well as important textual variables. The overall objective is to deliver a medical device classified as 2b, with an expected achievement of TRL7 by the end of the project. It is important to note that the assessment process will begin before the complete development of the AI-based tools. This requires an interpretation of the results from our Delphi survey with a dual perspective – considering both early and comprehensive assessments. This nuanced approach aligns with the changing nature of AI development and emphasizes the need for adaptable HTA methodologies to suit evolving technological landscapes.
Limitations and Developments
The authors acknowledge the research limitations at the current stage.
Regarding the Delphi survey, certain results may be influenced by the composition of the panel, which was not evenly distributed across areas of expertise, especially in the field of AI. Although we incorporated various perspectives, only a few patients/advocates chose to contribute to the study. To address this, the Patient Advocacy Lab of ALTEMS (Graduate School of Health Economics and Management) (Reference Petrangolini, Morandi and Di Brino44) will collaborate closely with the HTA group within the project to better capture patient perspectives. In the future, it should be considered the involvement of other categories of stakeholders (i.e., consumer and data protection organizations).
Additionally, no subgroup analysis has been conducted yet due to the limited sample size. We presented the list of candidate topics with a brief description (see Supplementary File 2) to our experts. Some topics relied on the EUnetHTA definition or interpretations found in studies. In other cases, we provided examples. Despite our efforts, it became evident before and after the survey that some topics lack validated and shared definitions, as is the case with explainability (Reference Amann, Blasimme and Vayena24;Reference Amann, Vetter and Blomberg45), interpretability (Reference Farah, Murris and Borget46), and trustworthiness (Reference Thiebes, Lins and Sunyaev47). Similarly, certain results, such as the exclusion of environmental-related topics, may indicate a lack of experience in assessing the impact of the environment on health care rather than the lack of relevance of the matter. The evaluation of environmental consequences remains an evolving field in HTA (Reference Toolan, Walpole and Shah48;Reference Greenwood Dufour, Weeks and De Angelis49;Reference Marsh, Ganz and Hsu50) with unresolved issues (Reference Guirado-Fuentes, Abt-Sacks and Trujillo-Martín51).
Moreover, some results require further investigation, as seen in the low level of agreement on organizational impact (see Figure 4). This is particularly important because AI is anticipated to disrupt the organization of healthcare services (Reference Bellucci52;Reference Kupietzky53).
AI and HTA: Remit of the Study
The aim of our analysis was to contribute to the definition of an HTA–AI framework, rather than investigate the availability of evidence required by that framework. Previous studies, such as Farah (Reference Farah, Davaze-Schneider and Martin25) and Di Bidino (Reference Di Bidino, Piaggio and Andellini54), have shown that current AI studies are not enough to meet HTA requirements. These studies have emphasized the need to improve evidence collection and HTA processes to adequately address the unique characteristics of AIHTs. Furthermore, our work does not examine the requirements and implications at the regulatory level. Lastly, the relevance of how AI-driven evidence could support assessments (Reference Zemplényi, Tachkov, Balkanyi, Németh and Al55;Reference Tachkov, Zemplenyi and Kamusheva56) is beyond the scope of our analysis.
Conclusions
The development of an HTA framework should not only consider the characteristics of the specific category of technologies but also reflect the level of agreement among experts regarding what to assess. To facilitate the identification of an HTA framework for AI, a Delphi survey was conducted, involving 46 experts who selected 48 topics out of the 65 proposed. Not all of these topics are currently included in the EUnetHTA Core Model. The feedback collected from experts will play a crucial role in both defining the HTA framework and testing it with AI-based tools currently under development in the EU project AI-Mind. This project aims to support the early identification of dementia in patients with mild cognitive impairment.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0266462324000308.
Funding statement
The author(s) declare financial support was received for the research, authorship, and/or publication of this article. This project has received funding from the European Union’s Horizon 2020 research and innovation program under grant agreement no. 964220. This paper reflects only the authors’ view, and the Commission is not responsible for any use that may be made by the information it contains.
Competing interest
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Open access
