An online survey platform (Qualtrics XM, USA) was used for survey hosting and data collection. The survey questions were formulated after discussion within the research team (Table 1). The survey was divided into five sections: Section 1 compiled the participants’ demographic information (age, gender, years of clinical practice), Sections 2, 3 and 4 assessed the knowledge, attitudes and practices of the participants respectively, and Section 5 was left open-ended for additional comments. A combination of multiple-choice questions, multi-selection questions, questions utilising the Likert scale and open-ended questions were incorporated.

Sample size calculation was performed prior to survey administration, and a target sample size of 52 was calculated for a 95% confidence level and 8% margin of error.

The sample was planned to be obtained using convenience sampling. Prior to dissemination, a qualitative assessment of the survey was performed to ensure validity. Content validity was verified via a review of the survey questions by a panel of three experts in the field of OMS and AI. These experts assessed the questions and gave feedback on the relevance and comprehensiveness of the survey. Subsequently, face validity was verified via pilot testing conducted with 6 participants (i.e., 10% of the target sample size). This was done as part of survey refinement to ensure question clarity, comprehensibility and relevance. The internal consistency of Sections 2, 3 and 4 (Knowledge, Attitudes and Practices respectively) was assessed using Cronbach's alpha, with a value above 0.7 considered as an acceptable level of internal consistency.

The survey was distributed to potential participants via email through local professional organisations (the Association of Oral and Maxillofacial Surgeons Singapore, AOMSS) and academic institutions (NUS). The survey was kept open, with no password required to enter the survey. Participants were informed of the aims of the study, the study team members and the approximate duration required to complete the survey, and consent was obtained upon commencement of the survey. Participation was voluntary, and participants were given the option to withdraw at any point. All responses were confidential, and participants were not asked to provide identifiers. No incentives were provided to participants who completed the survey.

Data collected via the survey platform was exported on an Excel sheet. Apart from the survey responses, data on the number of unique users was tallied using IP checks and cookies to sieve out duplicate responses. Timestamps were also recorded to identify surveys that were completed in under 20 s and surveys which were not completed, which indicated an inaccurately filled-out survey that should be excluded.

Descriptive statistics (frequencies and percentages) were computed for each survey question. In addition to this, statistical analyses were performed to identify any associations between demographic factors (i.e., gender, age, years of practice) and survey responses. For questions following the Likert scale, positive responses (“Strongly Agree” and “Somewhat Agree”) were combined to form one variable “Agree”, while negative responses (“Strongly Disagree” and “Somewhat Disagree”) were combined to form one variable “Disagree”. Questions that followed a five-point Likert scale had an additional response option “Unsure”, which was included as a third variable. The Fisher's exact test was then used to assess for associations between the responses and demographic factors. Similarly, for questions allowing multiple responses, each response option was treated as a separate binary variable (selected vs. not selected), and the Fisher's exact test was then performed to determine whether the proportion of each option differed between any demographic groups. All statistical analyses were conducted using the R Statistical Software, with the significance level set at p < 0.05.

Lastly, thematic analysis of the free-text responses to the open-ended questions in Section 4 and Section 5 was performed by members of the study team in phases, with data familiarisation followed by coding. The coded data were subsequently segregated into themes as deemed appropriate to explore the various themes of how to better integrate AI into practice.

Overall, the majority of participants reported lacking knowledge in the field of AI in OMS (Table 2). 60.4% of respondents answered “Strongly Disagree” or “Somewhat Disagree” to having a good level of understanding of AI in general, 52.1% responded similarly to being aware of the uses of AI in OMS. Only 18.8% reported having attended some form of training related to AI. There was no significant difference in responses by gender (p = 0.741–1.000), age (p = 0.153–1.000) or number of years of practice (p = 0.222–1.000) (Table 3). The Cronbach's alpha of 0.76 for the knowledge section indicated an acceptable level of consistency (Table 4).

When asked to list currently available AI technologies, 25 (52.1%) participants were able to list at least one related to OMS, of which 15 responses were related to examination and diagnosis, 7 were related to treatment planning, and 3 were for other uses. Outside of OMS, 24 (50.0%) participants could list a large language model, 5 (10.4%) mentioned robotics, and 5 (10.4%) mentioned speech-to-text and text-to-speech. 12 (25.0%) participants were unable to list any AI technology at all.

Most participants were found to have positive attitudes about the use of AI in OMS (Table 2). The majority of respondents agreed that AI can enhance patient outcomes (75.0%), that AI should be integrated into clinical practice (72.9%), and that AI should be a part of OMS training (68.8%), while only a small minority of respondents (4.2-12.5%) answered “Strongly Disagree” or “Somewhat Disagree” to the same statements. On the other hand, 79.2% of participants disagreed that AI may replace surgeons in the future (79.2%), while 58.3% agreed that overuse of AI may cause loss of clinical skills. While there was no significant difference in responses by gender (p = 0.078–1.000), a significantly higher portion of respondents aged 40 and below agreed that AI can enhance patient outcomes (p = 0.004) and that AI should be integrated into clinical practice (p = 0.009) (Table 3). Similarly, significant differences in the proportion of “Agree” responses were found between participants with 5 or less, 6–15, and more than 15 years of clinical practice (p < 0.001 and p = 0.024 respectively). The Cronbach's alpha of 0.71 for the attitudes section indicated an acceptable level of consistency (Table 4).

Regarding the advantages and potential concerns of AI in OMS, 83.3% of respondents reported the advantage of increased efficiency, 72.9% reported reduced workload, 45.8% reported increased personalisation, and 6.3% reported AI did not have any advantages. While responses were similar across most demographic groups, a significantly higher proportion of participants aged 40 or less found increased personalisation to be an advantage (p = 0.013), and the proportion of participants who found AI to increase efficiency significantly differed by their number of years of clinical practice (p = 0.010) (Table 5). 77.1% of participants reported concerns of inaccurate diagnoses or treatment plans with the use of AI, 70.8% reported concerns of over-reliance, 41.7% reported privacy and security concerns, and 41.7% had concerns of increased healthcare costs. These findings were consistent across all demographic groups (Table 5).

Although most participants reported having used AI outside of work (68.8%), only 25.0% reported having used it within the field of OMS. While 62.5% expressed that AI can make the completion of their work easier, most reported that they felt inadequately trained to do so (79.2%), and that their clinic was not equipped to do so (58.3%) (Table 2). There was no significant difference in responses by gender (p = 0.437-1.000), age (p = 0.290–0.880) or number of years of practice (p = 0.075–0.905) (Table 3). The Cronbach's alpha of 0.78 for the practices section indicated an acceptable level of consistency (Table 4).

Of the potential uses of AI in OMS, respondents stated that they had used or considered using AI most frequently for diagnosis (45.8%), followed by patient or student education (41.7%), self-directed learning (37.5%), treatment planning (37.5%), and as an intraoperative aid (16.7%). Responses were only significantly different between participants with different years of clinical practice (p = 0.020) in their consideration of use of AI for diagnosis (Table 5). 11 (22.9%) respondents stated they had never used nor considered using AI in practice; this was found in significantly higher proportion in clinicians with more than 15 years in practice (p = 0.009). The subspecialty of dentofacial deformities was considered for the integration of AI most frequently (50.0%), followed by implant surgery (39.6%) and surgical pathology (35.4%). No differences were found between preference of use in any subspecialty and demographic factors (p > 0.05).

For the open-ended question on what resources were needed for better integration of AI into OMS, four themes were identified. These themes are: (1) Access to training, (2) structuring, (3) funding, and (4) technological refinement. Ten (20.8%) responses fit into the first theme of access to training, with respondents stating that clinicians would benefit from structured training programs, courses, and access to relevant software. Five (10.4%) of respondents mentioned the second theme of the need for better structuring and workflows to facilitate smooth implementation. Under this theme, respondents mentioned the potential benefits of improved institutional and department workflows and sorting out of regulatory issues regarding consent and privacy prior to widespread implementation. For the third theme of funding, 5 (10.4%) of responses emphasised the need for better monetary support in the form of government or institutional grants to reduce the barriers to entry for integration. Lastly, 2 (4.2%) respondents mentioned the final theme of technological refinement. One respondent suggested the creation of a country-specific database to further refine AI models for better applicability to the local population, while the other mentioned their hesitance to use AI tools until the tools show better results.

Other valid concerns of the participants in this study can be divided into three broad groups: concerns regarding inaccurate diagnoses and treatment plans, privacy and security concerns, and increased healthcare costs. The four suggestions proposed by the participants to improve access to training, create better workflows for implementation, increase funding and further refine and optimise AI models are spot on in addressing the perceived issues revolving around using AI in OMS.

The possibility of inaccurate diagnoses and treatment plans was the most common concern among participants, with a small group even imploring more development and refinement of AI models before incorporating them into practice. The ability of AI to correctly answer OMS-related questions is still not ideal, with a study of large language models finding a mean score of just 62.5% or a B grade on OMS examination questions (19). However, emerging evidence shows the high diagnostic accuracy of AI models in detecting pathology using clinical records, photographs, radiographs or histopathological slides, and in the prognostication of oral diseases (20–23) Nevertheless, the repercussions of an incorrect diagnosis have led some to call for rigorous validation processes before implementation (24). Overall, overcoming this concern will require two improvements: the enhancement of AI models to optimise accuracy, and the spread of understanding that AI is not meant to replace the clinician, but to instead aid them in their tasks.

Concerns regarding security and a breach of patient privacy were expressed by almost half of the respondents. This concern is a major one, since large amounts of data are fed to and processed by AI models for their training and validation. Breaches in patient privacy during both model development and utilisation can occur, since acts of governance such as the Health Insurance Portability and Accountability Act (HIPAA) currently do not have clearly established guidelines covering AI-related technologies (25). Fears that large tech companies like Google can re-identify de-identified data through triangulation with other data sets are also not baseless, as re-identification has been proven successful before (26, 27), and lawsuits because of this have been filed (28). Resolving this issue necessitates two solutions. First, consideration can be made to develop future AI models with realistic synthetic patient data created by generative models instead of real patient data (29), to obviate the need for real patient data where possible. Second, regulations regarding patient data need to be revisited and refined to improve privacy legislation specific to the use of AI in healthcare.

On a similar note, although other ethical considerations of the integration of AI into OMS were not mentioned by any of our participants, it is a potential issue that needs discussion. As much as AI has the potential to improve patient outcomes, training and research, its implementation should still uphold transparency, informed consent of its use, and patient autonomy to disclose their health information and make decisions relating to their care (30). Professionalism must be maintained through a declaration of its use in practice and research to both patients and colleagues. Furthermore, it is paramount to understand that AI in its current stage should serve to complement, and not replace, the human touch still required in patient care, education and research. A review by Rokshad and colleagues highlighted a framework that can be implemented in the future refinement of AI models in dental practice and research; this framework addresses ethical challenges based on the eleven ethical pillars of transparency, diversity, wellness, respect of autonomy, privacy, accountability, equity, prudence, sustainability, solidarity and governance, and are a good guide to understanding how to best protect our patients' interests in the integration of AI (31).

Lastly, almost half of the respondents in this study cited the concern of potentially increased healthcare costs. Interestingly, this was in contrast to a similar study of nurses, where most participants thought that AI could instead reduce healthcare costs (16). While fears that the cost of development of AI technologies may be carried forward to the patient's bill, a review of 200 studies in 2022 has instead shown that integrating AI into healthcare reaps significant cost savings (32). This can be attributed to reduced diagnostic and treatment time and improved efficiency that increases with the number of years of its use. While no cost-effectiveness studies specific to OMS have been published, the abovementioned cost savings have been reported in other aspects of dentistry, such as its use for dental caries detection or the early identification of oral mucosal lesions (33, 34). To prevent increased individual patient costs, it may be in the interest of governments and institutions to provide funding for the development of these models. Other ways to reduce healthcare costs include pruning of the models to eliminate unnecessary components, as well as developing explainable AI models which incorporate a feedback loop to improve their usability and long-term sustainability (32).

This cross-sectional study is the first one conducted to evaluate the knowledge, attitudes and practices of specialists and trainees in the field of OMS, and while similar views are shared with other healthcare workers, gaining an understanding of the attitudes and concerns of people in our fraternity allows us to form concrete solutions to outstanding issues. However, the study has its limitations. First, due to the small population size of OMS in Singapore, the study may be underpowered. While 55 responses were initially recorded, the ultimate sample size was below our target sample size of 52, as 7 responses were incomplete and had to be excluded. As there is only one postgraduate training program in the country, the mindsets of trainees could also be skewed, thus potentially limiting the applicability of their perspectives to other populations. Consequently, the small sample size may imply results that are less generalisable and robust. Addressing this limitation will require multi-centre execution across various countries in the region to increase future sample sizes.

Convenience sampling was used for this study due to the limited manpower available in this project to perform more sophisticated sample methods (e.g., systematic or stratified sampling). While convenience sampling is simple to perform, it may introduce bias towards people who use technologies like social media more often and skew the results to favour technologies like AI. Fortunately, the small OMS population size and high digital literacy in Singapore meant that dissemination via emails and social media channels of local professional organisations was likely to reach the vast majority of OMS clinicians in the country.

Furthermore, the views reported in this study are specific to our population of mostly ethnically Southern Chinese clinicians, and may reflect different levels of acceptance towards AI compared to Caucasian, African or other ethnic groups (35). Finally, it is important to acknowledge that the results of this study only reflect the current population of clinicians in OMS. The views of clinicians may be very different ten years from now, due to the likely further advancement of AI systems and possibly legislative reforms and institutional shifts towards AI-driven care that can make its acceptance more widespread.