Artificial intelligence (AI) was first introduced some years ago, but in recent years, there has been increasing exploration of the utility and cost-saving of such technology (1, 2). AI brings about great potential in changing existing healthcare practice, from prevention, screening, diagnosis, treatment, and care (2, 3). AI could tap onto data from existing medical records; or even data from smartphones that individuals possess, and data from the applications that individuals use, including their social media posts (3, 4). By using large datasets and employ advanced techniques such as machine learning and deep learning approaches, AI informs more precise predictions of behavioral patterns and understanding of existent medical conditions (3). These benefits would facilitate the clinical decision process, improve the efficacy and accuracy of diagnosis, and diminish physician's workload. Evidence on the utility of AI in healthcare has been widely recorded from dentistry (5), primary care (6), radiology (7), ophthalmology (8) or pathology (9). AI has been recommended for inclusion in routine workflow processes (7). It is thus evident from these studies that the use of AI has been explored in various domains, and it is a promising technology for healthcare.

However, although many reports show the promising role of AI, the usage of AI is still in the early stage. Recent studies indicated low rates of physicians who were familiar or had chances to adopt AI in their clinical practices, even in technologically advanced nations such as 5.9% in South Korea (10) or 23% in the United States (11). Many technological, social, organizational, and individual challenges to apply AI principles in healthcare facilities have been discussed thoroughly in literature (2, 12–15). Nonetheless, the most important factor was physician's attitudes and perceptions toward AI, which can decide whether they would want to integrate AI in their practice or not (13, 14, 16). In healthcare, when the clinical decision is closely related to the patient's lives, health professionals are more likely to be cautious to use new technology in treatment and care; thus, it is not easy for them to trust and use a new product to support their practice.

Health systems can actively involve in the roles of AI adopters and innovators. Therefore, given the rapid expansion of AI applications in healthcare, it is crucial for future health workforces to prepare their capacities, as well as positive perceptions and attitudes to participate in the development of these novel tools. Prior studies indicated some controversial results about the attitudes and intentions to use AI in healthcare practices among medical students. For example, a study in the United States revealed that although the majority of radiology students had a belief in the future role of AI, they felt less interest in the applications of AI in the radiology field (17). Another study in the United Kingdom showed that 49% of medical students were more likely to apply for a radiology career due to AI (18). Understanding determinants of their behavioral intention to use and adopt AI in healthcare delivery is thus necessary for developing medical education curriculum to facilitate AI competence.

In Vietnam, it was been reported in 2019 that more organizations (both healthcare and non-healthcare related) have started developing AI technologies, and utilizing such technologies (19). In 2019, the Vietnam Ministry of Health has issued Decision No. 4888/QD-BYT about the applications and development of smart health care during 2019–2025, which underlines the importance of digital health and strategies to integrate digital health, including AI, into routine health service delivery (20). To date, there remains limited evaluation of AI amongst Vietnamese healthcare services. From our knowledge, there has been only prior publication, that of Vuong et al. (21) that presents a framework seeking to evaluate the AI readiness of the Vietnamese healthcare sector. The authors reported that the implementation of AI in healthcare in Vietnam is limited by several factors, such as the lack of funding; the necessary information infrastructure; and most importantly, the lack of understanding and misunderstanding of AI. Whilst the previous article by Vuong et al. (21) has provided some insights into the challenges with AI implementation and utilization, the review focused on issues at a macro-level, and has not evaluated the perspectives of individual healthcare professionals. For there to be a high uptake rate of AI on the ground, there needs to be an understanding of existing attitudes, preferences, and perspectives of future physicians.

In healthcare, various theories have been used to understand comprehensive facilitating factors in the individual's adoption and acceptance of a novel technology. For instance, several theories included the theory of planned behavior, the theory of diffusion of innovations, the technology acceptance model, or the unified theory of user acceptance of technology (UTAUT). Of which, UTAUT has been recognized as one of the most common theories to examine the adoption behavior of one individual (22–25). UTAUT was developed based on other dominant behavioral theories. Venkatesh et al. showed a higher explanatory level of UTAUT compared to other theories in exploring the information technology adaptation, with 70% of the variance for behavioral intentions and 50% of the variance for actual use (26, 27). A previous study in Chinese physicians showed that initial trust and performance expectancy were significant predictors for the AI adoption intentions (28). This study aimed to use UTAUT to explore the behavioral intentions of medical students to adopt an AI-based Diagnosis Support System. Understanding medical student's attitudes and perspectives would help to resolve potential barriers in adoption at the ground level, and such a survey would also help guide AI policy formulation at different levels.

In this section, we presented literature review and conceptual framework of this study. Moreover, study design, data collection method, and statistical analysis were described.

Table 1 depicts the demographic characteristics of our sample. The mean age of selected medical students was 20.6 years old (SD = 1.5). The majority of them were female at 73.5%, lived in urban areas (89.1%), and Ho Chi Minh city (59.7%). Most of the respondents belonged to the general physician program (57.8%).

Table 2 showed that the initial trust construct had the lowest mean score at 3.0 (SD = 0.9), while TC had the highest mean score at 3.8 (SD = 0.9). Overall, the Cronbach's alpha of each construct ranged from 0.738 to 0.909, suggesting good reliability among constructs. All item loadings of these constructs were above 0.7, and all construct's AVE values were above 0.5, indicating good convergent validity.

In Table 3, regarding discriminant validity, the value of the square root of AVE per construct was higher than its correlation coefficient with other constructs. Moreover, the results of VIF analysis showed that all VIF values were below 10, suggesting no multicollinearity existed.

Figure 2 illustrates path coefficients and p-values of PLS analysis. Regarding the behavioral intention model, only social influence (β = 0.527, p < 0.05) was positively related to the behavioral intention. Meanwhile, other constructs such as performance expectancy, effort expectancy, initial trust, and perceived substitution crisis showed no associations with behavioral intentions to use AI. Overall, the model with five proposed constructs for behavioral intentions, including performance expectancy, effort expectancy, social influence, perceived substitution crisis, and initial trust, explained 47.6% (R² = 0.476) of the behavioral intention's variance.

Figure 2 also shows that effort expectancy (β = 0.201, p < 0.05) and social influence (β = 0.574, p < 0.05) were positively associated with initial trust, while no association was found between performance expectancy and initial trust (p > 0.05). The model including performance expectancy, effort expectancy, and social influence explained 47.9% of the variance of initial trust (R² = 0.479).

Developing and adopting AI in healthcare are essential due to its great benefits in enhancing healthcare professional's performance and efficiency. Overall, the perceptions of our students about diagnosis-related capacities of AI, effort to use AI, and intention to use AI were positive. It is clear that the role of AI in healthcare delivery has been widely documented, where AI has shown its success in the interpretation of image and examination data, as well as clinical outcomes prediction and management (15, 51). Nonetheless, information about AI and its application in Vietnam has been disseminated in mainstream media but not in university settings. Our results indicated that undergraduate medical students in Vietnam had great confidence in the knowledge of their work characteristics, understanding how AI could assist them to promote diagnosis performance, and desire to use AI when available. However, there were still some gaps between their expectancy and preparation, including awareness of technology characteristics and capacities to use such technology. Equipping the medical students with the basics, as well as the correct understanding and attitudes about the application of AI in medicine, are crucial in the digitalization of the healthcare system. However, currently, the medical training program in Vietnam has not been systematically updated in this area. The AI content has been mainly shared through scientific seminars or short-term courses, without a specific program to develop the AI capabilities.

This lack of pre-paration might also lead to the findings that the majority of our sample somewhat agreed or agreed that AI would replace the position of physicians in healthcare. This result was congruent with findings among medical students worldwide, particularly those in the radiology field (13, 14, 18, 52). Several previous studies found contradict results where the medical students stated that AI could not have a role as an alternative for the physicians in the future (33, 53), particularly in some fields that need a “sense of caring” or “art of caring” such as psychological health or aging care (53–55). Many authors argued that AI should be treated as a virtual assistant rather than being a replacement for physicians in healthcare. However, prospective physicians should acquire fundamental knowledge about mathematics, data science, AI, as well as ethical and legal issues related to AI (56). They should understand the systemic bias behind AI algorithms due to the insufficient data, which might be a great reason for health equity issues when making a clinical decision (57, 58). Moreover, other humanistic aspects such as communication skills, empathy, decision-making, or leadership skills should also be required (53). Acquiring these capacities would enable physicians to take advantage of AI in integrating it into their routine clinical practices. Thus, it is needed to call actions to innovate the medical education programs in the digital area.

Our path analysis showed the dominance of social influence on the intention of using AI for future work among undergraduate medical students, instead of other factors such as performance expectancy or initial trust, which were found in the previous research (28). Although this result is unexpected compared to what we hypothesized, there were several reasons which can be used to explain this phenomenon. First, this study was conducted on undergraduate medical students, whose healthcare delivery experience, as well as perceptions about the diagnosis process, were constrained. Moreover, given that AI has not yet been scaled up in Vietnam and AI-related curriculums for medical students had not yet been developed, we supposed that the majority of our sample had no experience with an AI-based diagnosis support system. This limitation hinders the way medical students perceived their capacities in adopting AI, as well as results in the homogeneity in their competency and trust evaluation. Moreover, because of this lacking experience, it is understandable when undergraduate medical students tended to be heavily dependent on the experiences of senior physicians in their social networks and information they gathered in social media about AI. With the exchange and sharing of practical experiences from those who have used this AI system, students' trust and intention to use the AI system in the future would be improved.

The findings of this study suggested several implications. First, undergraduate medical students should actively find opportunities to update and involve in AI development and adoption to increase their necessary AI knowledge and capacities. Self-learning ability is important to acquire new knowledge in the context where AI curricula at medical schools have not been paid sufficiently. Second, our study suggested the importance of role model approaches for facilitating the use of AI in this group. Opportunities to gain hands-on experience in different teaching hospitals are critical. AI may be useful for diagnosing rare conditions, which are often only seen at large teaching hospitals. Finally, this study underlined the need to integrate AI curriculums in the current medical education, which helped medical students to prepare appropriate capacities in technology adoption. Further studies should be performed to measure the preference and effectiveness of different education strategies to facilitate AI applications in healthcare among health professionals and medical students. Moreover, they should also assess whether training students with AI helps or hinders their diagnostic abilities.

Some limitations should be acknowledged in this study. First, since our study was conducted on medical students who had no experience with AI-based diagnostic support systems, we have not yet assessed whether they would use these systems or not in the future. A longitudinal follow-up study evaluating the rate of use of this system among medical students after graduation is essential to help refine the theoretical model. Second, our research was conducted online and had recruited medical students in entire Vietnam; however, this study may be limited to the group of medical students with Internet access, while other groups of medical students were not accessed. In addition, a small sample size might reduce the statistical power. Other studies on larger sample sizes need to be conducted, which help verify our results in other medical student groups. Third, in addition to constructs included in the theoretical model, the study has not assessed the mediating effects of other factors such as age, gender, and previous training in AI use during university studies, etc., which could affect the relationship among factors in the theoretical model.

This study highlights positive behavioral intentions in using an AI-based diagnosis support system among prospective Vietnamese physicians, as well as the effect of social influence on this choice. The development of AI-based competent curricula should be considered when reforming medical education in Vietnam.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Institutional Review Board of Youth Research Institute. The patients/participants provided their written informed consent to participate in this study.

LN: conception, design, acquisition, interpretation of data, drafting the article, and final approval of the version to be published. HN: conception, design, drafting the article, and final approval of the version to be published. AT: design, analysis, interpretation of the data, drafting the article, and final approval of the version to be published. CN: acquisition, analysis, interpretation of data, drafting the article, and final approval of the version to be published. LV and MZ: conception and design of data, revising article, and final approval of the version to be published. CH: conception and design of data, drafting the article, and final approval of the version to be published. BT, CL, and RH: conception, design, interpretation of data, drafting the article, and final approval of the version to be published. All authors contributed to the article and approved the submitted version.

This study was funded by NUS iHeathtech Other Operating Expenses (R-722-000-004-731) and NUS Department of Psychological Medicine Other Operating Expenses (R-177-000-003-001).

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.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.