Study design and target population

A cross-sectional study was carried out in Slovenia and Croatia in 2023. An online survey was conducted among researchers and teachers at medical schools. The link with the invitation to participate in the study was sent to all researchers and teachers employed in the departments and institutes at two Slovenian and four Croatian Faculties or Schools of Medicine (Faculty of Medicine, University of Ljubljana; Faculty of Medicine, University of Maribor; School of Medicine, University of Zagreb; School of Medicine, University of Split; Faculty of Medicine, University of Rijeka; Faculty of Medicine, University of Osijek). Two reminders were sent to facilitate the participation in the study and increase the response rate. In Slovenia the data was collected between 15 May 2023 and 12 June 2023 in Ljubljana and between 12 October and 7 December 2023 in Maribor, and in Croatia between 16 October and 27 November 2023. The link with the invitation to participate in the study was sent through the dean’s office via e-mail to 1764 (Slovenia 536, Croatia 1228) RTs which were employed at the time. Ethical approval was obtained from the Medical Ethics Committee of Republic Slovenia (no. of approval: 0120-87/2023/6; October 5, 2023) and from ethical commissions of 4 faculties of medicine in Croatia (Zagreb no. 641-01/23-01/01; September 20, 2023; Split no. 2181-198-03-04-23-0069; September 27, 2023; Rijeka no. 2170-1-42-04-36/1-23-7; September 29, 2023; Osijek 602-06/23-08/03; September 18, 2023). The participation in the study was voluntary and anonymous–researchers willing to participate in the study followed the invitation link and by doing so agreed to participate in the anonymous study. Additional written informed consent was deemed unnecessary.

Sampling method

The study aimed to include all RTs employed in the departments and institutes at two Slovenian and four Croatian National Schools of Medicine, as survey participants. The link with the invitation to participate in the study was sent via e-mail to 1764 (Slovenia 536, Croatia 1228) RTs which were employed at the time.

Data collection tool

From the authors Martinho et al. [25] we obtained written permission on December 5, 2022, to translate, adopt and use their questionnaire on ethics surrounding health AI in our study (S1 Appendix). The questionnaire included 40 items to which respondents replied on a five-point Likert scale of agreement (from completely disagree to completely agree).

The questionnaire was localized to Slovenian and Croatian environment [27]. Two independent translators translated the questionnaire into Slovenian and Croatian language, compared the statements and resolved the discrepancies. After reaching consensus, the questionnaire was back-translated to English and the meaning of the items was investigated. The final version of the questionnaire was obtained. Cognitive interviews with 28 researchers, teachers and/or physicians were performed to carefully examine the meaning of each statement and further improve the understanding of the statements and evoke the same interpretation of the statement in all respondents. The extra statement was added to determine the preliminary status of AI in medicine (“Artificial intelligence is influencing decision-making in clinical and preclinical medicine”). The section on demographic information included age, gender, length of professional experience, and field of activity, where they could choose from options (a) clinical practice, (b) preclinical practice, and (c) other. In the invitation letter we specifically asked addressees to participate in the survey on the ethical aspects of the use of AI in medicine and healthcare. Final questionnaires, used in Slovenia and Croatia are provided in S2 Appendix and S3 Appendix, respectively.

Statistics

Categorical variables were described by frequencies and percentages, non-normally distributed continuous variables by medians and interquartile ranges. All respondents did not provide answer to all questions–all the answers provided were included in the analysis. Comparison between the two countries in the sample characteristics were performed using chi-square test for categorical and Mann-Whitney U test for continuous variables. Prior the analysis, values of each statement, measured on a five-point Likert scale of agreement, were recoded into two categories. The first included the answers from 1 to 3 indicating lower agreement with the statement and the second the answers 4 and 5 indicating higher agreement with the statement. The association between each recoded statement and country was examined using univariate logistic regression analysis.

To examine the perspectives of researchers and teachers, principal component analysis with the orthogonal rotation was performed on the original items about the use of AI in medicine. Principal component analysis reduces the number of variables into components–linear combinations of the measured variables that maximally explain the variance of the variables. It is used to identify patterns in the data. To distinguish between perceptions of the RTs in the two countries, principal component analysis was performed separately for each country. The number of components extracted was based on the examination of the scree diagram [28], the number of items with high loading on each component and component interpretability. Furthermore, the broken stick method [29] was considered in which the eigenvalue of a given component is compared to the value of the one that would be expected by chance (if the total variance would be distributed equally among the components). Components with higher eigenvalue after varimax rotation than generated by the broken stick model were retained. The names of each component were based on the original research of Martinho et al. [25]. In their research they predefined ethic clusters “compiled from 22 major guidelines of AI ethics as a guidance tool” (pp.3). Items of the questionnaire (S1 Appendix) were assigned to clusters Privacy (statements 1–4), Fairness (5–8); Accountability (9–10,40); Transparency (11); Safety and cybersecurity (12–13,39); Human Oversight (18); Explainability (15–17); Future of Employment (19–20, 22); Responsible Research Funding (23–24); Education About AI (25,34); Human Autonomy (18); Certification of AI products (29–30); Ethical Design (31− 33); Health specific deliberations (14,21,26–27,36–38); and AI in the Covid-19 pandemic (28,35). Our components were named after the most prevalent statements from Martinho et al. [25] clusters with highest weights on each component. All statistical testing was performed at the significance level α = 0.05. No correction for multiple testing was applied. Program IBM SPSS, version 28 was used for the statistical analysis.

Ethical design perspective

Ethical principles are, since Nurnberg Code, the foundation for research and experiments where human beings are involved [30]. Moreover, it is imperative that such principles are consistently integrated within the context of new technological advancements, notably AI, as well [17, 31]. In that context our ethical design perspective confirms that statements on ethical dilemmas with the highest share and agreement per country (above 85%), analyzed with univariate logistic regression, did not show any statistical differences between two countries on two of the three statements relating to the ethical issues. The statements highlighted the importance of ethical principles that must be in line not only for RTs but also for AI developers. Croatian RTs focused more on AI developers to be bound by medical ethics, while Slovenian RTs focused more on attitude that AI healthcare products should be tested in randomized controlled trials (RCTs) which is in line with the Thomas Grote’s view that the clinical benefit of AI systems needs to be studied in clinical trials-particularly RCTs [18]. In both countries RTs strongly stressed that AI must be in line with ethics, an attitude that was recently discussed thoroughly in an article of Jeyarman et al. [31] and other recent studies [18, 32, 33]. The results align with those of Marthino et al., but they identify an additional attitude of privacy as a core ethical value, which is not evident from our perspective [25].

Health-specific deliberation perspective

Statements with the highest agreement per country analyzed with univariate logistic regression showed significant differences in two of three statements in this perspective. Slovenian RTs agree to lower extent than Croatian RTs that most areas of healthcare can benefit from the AI use and that RTs could focus on more complex tasks because of the AI, while there are no differences in the statement that AI will enhance medical decision making in situations of care rationing. Despite lower agreement among Slovenian RTs those three statements are highly consistent with each other in both countries i. e. that AI will enhance medical decision making in situations of care rationing, allow providers, clinicians and staff to focus on more top-of license skill sets and activities together with believe that most area of healthcare can benefit from AI.

The three statements in health specific deliberation perspective emphasize again what are the expectations of RTs in Croatia and Slovenia as worldwide. The study by Katirai et al. [20] discussed the perspectives of patients and the public in Japan regarding the use of AI in medicine. They conducted a workshop, leading to the identification of 55 expectations and 52 concerns related to AI, which were categorized into 12 major themes. They found expectations for improvement of hospital administration, quality of care and patient experience, and positive changes in roles and relationships, and reductions in costs and disparities. However, were concerned about healthcare changes, loss of autonomy, risks of accountability, data management and disparities. Comparably in empirical study by Amann J. et al. [34] participants highlighted, besides all technical benefits of AI, the significance of relational aspects, expressing concerns about how AI influences roles, responsibilities, and the rights of patients to receive information and participate in decision-making processes. Moreover, not only ethical aspects but also sociological aspects are important in implementation of AI in healthcare which is comparable with the robotics care for older people [35]. These views are comprehensible since current medical AI systems do not have a capability of autonomous decision-making which thereby disqualify them as responsible agents. Consequently, the responsibility for AI’s actions should be firmly put on the human agent [36]. Our perspective aligns with the first one by Marthino et al. [25], which emphasized that AI is a useful tool allowing doctors to focus on top-of-license skills. However, they stress that, despite the positive outlook, medical doctors must remain in charge of both the medical decision process and the design process. These two attitudes are touched upon in our other perspectives, with no difference in results.

Explainability and transparency perspective

Transparency and explainability are crucial and increasingly discussed attributes of AI systems, yet practical guidelines for defining these requirements are still subject of discussion for these are abstract terms that have a very concrete application [37]. Drobotowicz K. et al. in their empirical study extracted the following parameters for transparency: knowability of purposes, disclosing the nature and sources of the data they are utilized, providing easy explanations for users to understand, obtain individual insight into data management, human involvement in AI services [38]. In our third perspective statements with the highest or lowest agreement per country analyzed with univariate logistic regression showed significant differences in two of the three statements, i.e., only reliability of AI is of interest of RTs and appropriate informed consent is not possible while no differences were found in statement that AI is only used by clinicians that understand how AI decisions are made. The share of respondents that agree or strongly agree with the statements higher than 60% in both countries showed that Slovenian RTs agree more with the statement that appropriate informed consent is not possible. The share of respondents that agree or strongly agree with the statements lower than 60% in both countries showed that Croatian RTs to higher extent than Slovenian RTs agree or strongly agree that only reliability is of interest to RTs.

However, those three statements have rather diametrically opposite component weights in both countries. While high positive weight among Croatian RTs with variable “healthcare professionals do not need to know how AI medical tools work, only reliability of AI is interest of RTs” and high negative weight for variable “appropriate informed consent is not possible” diametrically opposed weights are obtained among Slovenian RTs.

These diametrically opposed meaning cannot be easily understood and put in the right context if we do not look at the second statement “appropriate informed consent is not possible” where diametrically opposed meaning of the statement is found again. Croatian RTs are sure that they can get appropriate informed consent because in their opinion there is no need to know how AI medical tools works. Slovenian RTs cannot explain how AI medical works and therefore appropriate informed consent is not possible. The third statement, indicating a Croatian RTs attitude with the component weight of -0.58 and a Slovenian RTs attitude of 0.61, supports the first two statements in both countries. This shows a consistency in the attitudes of RTs from both countries regarding the first statement. In other words, in Slovenia it is stressed that RTs need to know how AI works, how AI medical tools work, and without these two statements appropriate informed consent is not possible for them. In Croatia is stressed that RTs don’t need to know how AI works, how AI medical tools work, and because of that they believe that appropriate informed consent is possible for them. Numerous articles discuss the ethical challenge of explicability in AI, especially in medicine. In the context of our study, it is useful to highlight Adams’ proposal to include a ’principle of explicability’ alongside the traditional bioethics principles, and argues for explicability as an essential ethical consideration, enabling a bridge between technical demands and high-level ethical standards in AI use [11, 34, 39]. In their fourth perspective, Marthino et al. link explainability with education, while in our case, it is connected with the need for transparency. They state that explainability is a key value in the sense that, in order to reap the benefits of AI, physicians must understand and lead AI technological progress. They stress that physicians are interested in learning about AI and have time to learn how to use complex AI-based medical devices [25].

Privacy perspective

The following perspectives differ in name and statements. Another concern expressed by RTs in both countries relates to Privacy. The use of AI in medicine is challenged by protection of data privacy as data can be impacted by manipulative goals [19, 40]. Privacy and confidentiality are the single most important statements with the highest weights per component seen only among Slovenian RTs and not among Croatian RTs, where privacy was joined with autonomy. In Slovenia RTs believes that confidentiality should not be the highest priority in AI-based healthcare or constrain the implementation of AI but that it will have to be redefined. Privacy is defined through confidentiality. In Croatia, however, confidentiality is closely connected with decrease of the autonomy of the RTs due to use of AI. They believe that it will decrease the autonomy, reshape the patient-physician relationship and therefore influence the very meaning of confidentiality. For comparison, in Marthino et. al. privacy is considered a core ethical and medical value in the perspective of ethics, i.e., it is linked with the ethical demands for AI implementation. From this perspective, AI will not increase discrimination and improving equity and inclusion is not mandated to be priority [25].

It is not clear why both statements have not so high weights in Slovenia, but it could be hypothesized. Firstly, that those applying AI in their work found that sharing data with AI medical systems cannot protect individual privacy and confidentiality. In that line confidentiality must be redefined. Secondly, according to paradigm of the evidence-based medicine, RTs involved in decision making process are not bound with classically patient-physician relationship. Thirdly, development of AI found RTs unprepared towards AI medical companies or they are excited about possibilities which AI offer while forgetting privacy and confidentiality. Indeed, the significance of these hypotheses is in line with the literature available [21, 22, 41].

Autonomy and privacy/ fairness and autonomy perspective

RTs in Slovenia linked the autonomy with one of the core medical ethics principles–fairness (justice). They consider that the use of AI will change medical practice, and the ethical principle of fairness ensuring medical decisions are made fairly could be threatened. One of the major concerns expressed by Slovenian RTs are that AI will increase discrimination and that the mantra “fail fast and fix it later” will put patients at risk. In both countries RTs believe that AI will decrease autonomy and essentially change patient-physician relationship. In Croatia they think that confidentiality will change too, while in Slovenia more emphasis is put on concern about the increase of discrimination and risk for patient. These concerns are recognized in recent literature as well. Vearrier et al. [7] maintain that this relationship must be viewed as a physician-patient-machine (AI system) relationship. Despite all the benefits of AI systems, the physician’s roll is crucial for e.g. interpretation of data in the clinical context of an individual patient. In line with that a physician must advocate for patient confidentiality. In his study Kiener [41] notice, that in accordance with the current clinical practices, there are specific situations where risks must be disclosed to the patient. Otherwise, patient’s informed consent will be compromised or it will be breached the broader duty to alert them about possible adverse outcomes, what Slovenian RTs in our study apparently recognized. The issue of discrimination is essentially a matter of contextual knowledge, which AI systems lack, hence this requires systematic and ongoing intervention by the human. The fundamental issue on AI system implementation is the uncertainty of discrimination due to bias [15]. In Marthino et al. it is shown that big companies are not to be trusted, especially regarding medical data ownership. Their results show that AI will not increase discrimination based on predicted future problems if privacy becomes a key ethical value. Additionally, they emphasize that physicians must remain in charge in the medical decision process [25].

Employment, safety and funding/ education about AI and safety perspective

The share of respondents agreeing or strongly agreeing with the statements regarding employment, safety and funding/education are well below 60% (lower than 40%, and down to 8%) in both countries. The countries also differed in perspectives. In the sixth perspective in Slovenia Education and safety RTs bring forward the lack of time or interest for learning about the AI tools on one hand and securing that big companies do not enter the AI medical market on the other addressed. There are similar concerns about the knowledge and skills of RTs regarding AI in similar surveys [23]. In Croatia one of the concerns is expressed in the fifth perspective on Employment, safety and funding. RTs expressed doubts that AI might not meet the expectations and that automate will not function in hospitals. It can lead to over-testing and over-diagnosis and hence worsening the problems in healthcare. As in Slovenia, also in Croatia the RTs discourage the entry of big companies in the health care space. Which can be considered justified since it concerns a relevant issue of liability and responsibility of RTs for AI systems failures [14, 28, 42]. Croatian RTs to higher extent agree that it is not difficult to operationalize clinical practice for a machine. Other authors such as Tahri et. al. [14] asserted that prior to undertaking such procedure, there is a necessary prerequisite for the humanization of AI, to guarantee that its design and implementation are in conformity with ethical principles. Nevertheless, these perspectives don’t seem to be very important for Croatian or Slovenian RTs. Marthino et al. emphasized that it is difficult to operationalize clinical practice for a machine and that physicians are interested in learning about AI. In their results, employment is linked with usefulness, showing that AI will not indiscriminately cause unemployment, and safety is associated with regulatory issues [25].

This study indicated the need for the further similar research, as it is the first in Slovenia and Croatia that comprehensively deals with the issue of RTs attitudes and perspectives on the AI use in medicine. These results provide a useful insight into the differences and similarities between the two neighboring countries. In spite the cultural and social similarities of these countries there are characteristic differences in views on AI use in medicine. There are some limitations of this study. It included only two countries, which is not enough if one wants to get a wider picture. Future research should involve RTs from a variety of countries to identify and compare differing attitudes and perspectives. In our research, we focused on RTs, i.e. highly educated health professionals who work in educational institutions, other profiles of health professionals are not covered in this study. This expanded scope would allow for a broader analysis of responses gathered from individuals within different kinds of medical institutions. Our findings could be of a great help in developing policies and legal regulations on AI use in medicine.