This study outlines EP work-activities that are the highest priority for new AI-tool development. Survey participants were asked to consider the development of a fully translated AI-tool for patient care that would be available at most EDs in Canada in the next 10-years. To triangulate responses, participants were asked to rank a list of common ED work-activities that may benefit from AI, and a list of existing AI-tool examples.
The survey sampled 5.65% of Canadian physicians practicing emergency medicine, not including residents, based on 2019 data from the RC(16) This estimate does not account for physicians practising EM with other licence types; for example, CCFP physicians without additional EM designations. Additionally, 6.07% of trainees were surveyed (33 of an estimated 543 active residents); based on the 2019 residency quotas.(16) In general, the breakdown of survey respondents fit with national trends. For example, responses by licence-type are similar to the RC reported proportions; however, this survey had slightly higher PEM representation.
Considering geography, there was a disproportionately high response from the Maritimes (29.2%); the author’s practice location being Halifax. The other highest response rates come from Ontario (29.2%), Alberta (11.4%) and Quebec (10.0%); this result may relate to each region having large AI institutes (Vector, amii, Mila, respectively) with provincial strategies for AI adoption. The results are also biased towards urban practitioners, with only 11.4% practicing in rural or regional centers; important input from rural physicians may been missed.
Concerning familiarity with technology in-general, respondents were neutral; approximately half neither “dislike” nor “like” technology and 9.0% indicated “no interest in technology.” The average TechPH index agrees with the finding that most respondents were neutral regarding technology interest.(17) Overall, these outcomes are reassuring that the respondents include general EP and are not necessarily biased towards physicians hyperspecialized in technology development, nor are they actively opposed to the integration of new technology.
Respondents have low overall experience with AI in their personal lives, clinical roles or work as EPs. We speculate that the ‘low’ personal experience with AI may relate to misconceptions about the technology, as we assume that most Canadians are daily consumers of AI-enabled apps and productivity tools (weather, navigation, search-engines, voice-to-text). This response may also be from the framing and interpretation of the question.
When asked about their understanding of AI-technology, 87.2% of respondents “agree” or “strongly-agree” they “understand what is meant by AI.” However, only 23.6% of respondents had a completely correct definition of AI (see Appendix D). These results suggest that more education around the concept and purpose of AI may be needed.
Few respondents have conducted any research in AI (4.5%). This result agrees with follow-up questions, where most respondents indicated “no experience at all” (71.4%) or “very little experience” (11.1%) with AI-research. Again, these findings corroborate the neutral TechPH index. However, almost all respondents either “somewhat agree” (39.8%) or “strongly agree” (48.5%) that they are interested in AI for EM.
Overall, EPs agree that AI has potential use for EM; however, physicians feel there will be only “slight” (60.6%) to “moderate” (35.9%) impact on how their job will change. This result suggests physicians believe AI will enhance current roles but not disrupt the specialty over the next 10 years. This opinion agrees with findings from surveys of psychiatrists and family-physicians.(9)
Considering EP impressions about AI’s capabilities, most thought AI was “likely” or “extremely likely” to be able to provide documentation. Interestingly, much of the current focus of AI development aims at tasks such as reaching a prognosis, formulating a treatment plan, formulating personalized medications and evaluating when to refer to a specialist, despite these being ranked either “neutral” or “likely.” Providing empathetic care was ranked as “extremely unlikely” or “unlikely”. These findings also match the opinions of psychiatrists and family-physicians surveyed with the same instrument.(9, 12)
Respondents also indicated examples where they “have used” or “heard of” AI being used in EM (Table 6). They were provided specific examples of AI-tools, and for triangulation, also general ‘work-activities’ where AI is used. Of note, the ranks of “have used” for ED work-activities, do not map to the AI-tool example ranks; the first choice AI-example ‘clinical prediction rules’ maps to the fourth choice AI-work-types ‘AI-tools for making the diagnosis, selecting investigations, etc.’. One explanation is that physicians may not agree on how to classify different types of AI-tools, or there are other more important AI-tools within the work-type categories not listed in the examples.
As well, interpreting Table 6 in the context of Table 7, many of the “have used” items fit into the “prognosis/diagnosis” and “formulating a treatment plan” categories, which are all areas that physicians have guarded opinions about. Interestingly, the ED-documentation-tools have only been used by 13.2% of respondents and only 41.1% had heard of AI-tool examples. Yet, this was the task with the best perception of being accomplished by AI. Additionally, all categories of AI-imaging interpretation were ranked low in terms of past use by EP; a surprising result given the large body of AI-research for radiology. Perhaps the use of this technology by Radiologists is not immediately obvious to the EP receiving the reports.
For the study’s primary outcome, there is clear consensus for translation of AI-tools to facilitate documentation, and as mentioned, respondents had the most confidence that AI could facilitate this task. Although many new ED information systems (EDIS) have some AI integration, as indicated, few respondents “have used” or “heard of” ‘automated charting or report generation’ and only 37% have “heard of” and 20.1% “have used” AI-tools for documentation. Based on the responses, we would suggest that tools for documentation are prioritized to meet both the expectation and needs in EM
The ‘documentation’ category is broad, including electronic charting with voice-to-text, or active listening with AI-powered scribes, or AI-powered summaries of patient records to consolidate them into succinct and accessible formats. Future work should clarify these needs in detail, perhaps using focused interviews.
Separate from these specific recommendations, this survey provides insights into a potential strategy for implementing AI tools in an ED setting. As such, we recommend that (i) ED physicians be engaged in the specification, design, evaluation, and implementation of future AI driven tools; (ii) Priority should be placed on developing proof of concept AI-solutions for the high-yield problems identified by ED physicians; (iii) Solutions should embed AI tools within the ED’s existing digital infrastructure and clinical workflow; and (iv) Developers should identify measurable and impactful outcomes for AI use, and use standardized metrics to assess these outcomes.
In conclusion, AI in an ED setting can be seen as an innovation agent, as the analysis of ED data can generate new insights about the effectiveness of certain procedures/policies and lead to the optimizations of ED resources. AI is not here to change ED practices, rather it offers solutions to optimize a number of practice challenges. The survey responses clearly point to perceived value for AI in the ED, however certain activities are more amendable to AI driven support. For instance, automated charting particularly using speech recognition and transcription, rapid interpretation of real-time ED data for clinical decision support, patient risk stratification, and forecasting for staffing. The opportunity to benefit from AI based applications relies heavily on their integration within the current clinical workflows and the data sources used by ED physicians. This will ensure that ED physicians do not need to change their practice to make use of AI tools, rather AI driven support is seamlessly available at the point of care. Overall, the growth of AI in medicine is on the rise and it is fair to conclude that the use of AI in ED is quite near in the future.