Artificial intelligence (AI)-powered tools are becoming increasingly common within clinical practice and medical education, and using AI in simulation as a pure learning tool and for assessment has been widely discussed. Targeted, efficient use of such tools has been shown to aid medical education and professional development [1].
Virtual patient educational tools and simulators have also demonstrated benefit for medical education [2]. Here we outline a hybrid approach which is for the educator to utilise AI to rapidly create an educational tool, whether that uses AI or not. We develop and evaluate the efficacy of an example simulator to illustrate the process.
The following sections will outline a practical five-step process to enable educators to identify use cases for learning tools or simulators, and then realise them using freely available AI coding aids:
- Identifying a target: what unmet need or gap could appropriately be addressed with a learning tool?
- Ideation: how could this be simulated? What specific aspect or skill is being tested?
- Creation: the fun begins!
- Iteration: how could the tool be improved?
- Evaluation: does it work? What do end-users think?
Identifying a target
The majority of the effort in designing a learning tool lies in identifying a suitable target and learning approach. Consider in your work or learning environment what focussed task, skill or niche could benefit from being addressed with a simulated eLearning tool. Once a need has been identified, consider whether a learning tool is really the best option. There are many ways to facilitate learning, and a model or tool is only one of them.
Ideas for modelling
You can begin by asking several questions, such as ‘How could this task or skill be simulated?’ and considering how to break down the skill into smaller parts (not dissimilar to the second of Peyton’s four steps [3]). Identifying which parts are the most important will aid in the design process later.
Recognise also that the fidelity of a simulator is not always the most important factor. It should be targeted at developing the specific aspects, skills or processes that require addressing. Research has shown that even though high-fidelity simulation is often associated with higher learner satisfaction, outcomes are comparable with well- designed, low-fidelity simulation [4]. In fact, high-fidelity simulation can be associated with learner overconfidence [5]!
An analogue example of a targeted-fidelity simulator is a venepuncture arm. These are commonly not attached to a ‘body’, but the simulated skin, texture and positioning are carefully designed to allow users to practice key venepuncture steps: (1) identifying a vein, (2) palpating the vein, (3) lining up the needle angle and approach, (4) inserting the needle to achieve flashback once appropriately sited. The targeted fidelity of the venepuncture arm is key – it is realistic and helpful in aspects of the target skill. Alternatively, if the task was at a communication station; explaining venepuncture to a patient and consenting them for it, there would be an entirely different set of fidelity requirements. One might not even need a venepuncture arm, just a model patient for the learner to talk to.
Figure 1: A flow chart demonstrating the process of carrying out artificial intelligence co-creation of learning models and simulators. Feedback is essential to refining the product, consistent assessment and iteration is key to this.
Using the AI aids
A plethora of AI coding aids and assistants are available, with some of the most common compared in Table 1. Most have some form of free plan available, and all feature a text prompt approach to produce code output. Some require a downloaded application or account creation processes to access fully.
Table 1: A table comparing three commonly used AI coding tools, summarising their sign-up, download requirements and request limits. Note GitHub’s CoPilot aid allows users to select from a variety of AI models (GPT, Claude, Grok, Gemini – some require paid subscription).
Precise prompting
All of the listed aids are powerful, but they are limited to the directions that you provide them. Forming a precise and descriptive prompt is key to achieving a good output.
Specificity is important in a prompt. It can help to outline some of the background before going into what you want the simulator or learning tool to do. For example, describing the level of the user as well as what the aim of the simulator is. This includes the visual appearance of the tool as well as how you want the user to interact with it. For example, using a slider, switch or typing in text. From there, you can consider how the simulator will provide feedback to the learner using it.
Figure 2: A screenshot of our RAPD simulator. This AI co-created learning tool aimed to increase user confidence and proficiency in testing and identifying RAPDs. Users are invited to swing the torch over the simulated eyes, and suggest whether an RAPD is present or not. Feedback is provided once the answer is submitted.
An example that we have tried is creating a pupil simulator to allow students and junior clinicians to practise testing for, and identifying, a relative afferent pupillary defect (RAPD), shown in Figure 2 and available freely online at: www.wkxm.org/rapid. Table 2 outlines how our simple framework for effective AI co-creation led to the creation of our RAPD simulator:
Table 2: A simple framework to consider how your target can be utilised to plan and develop an AI co-created learning tool.
Another example is a virtual optokinetic drum (OKNdrum) we created which is likewise now online and freely available (wkxm.org/vokndrum, shown in Figure 3).
Figure 3: A screenshot of our OKNdrum simulator. This AI co-created simulator aimed to provide an opportunity for medical students to learn about the optokinetic reflex, with additional (non-validated) use as a back-up clinical tool.
The requirement was for a backup clinical aid to show learners who might not have access to a niche tool, how it works. It is a simple set of black and lines in motion. The simulator allows the user to toggle line width, direction and speed. Learners, such as undergraduate medical students, can use this with peers to elicit an optokinetic nystagmus. In this case the tool does not provide any learner feedback, but is purely experiential. It is potentially useful for low-resource settings (being HTML based, it will run on any device with a browser), although it has not yet been validated for clinical use.
Iterating
Artificial intelligence coding aids can produce output that is wildly different to that anticipated, requiring iteration to continually refine the input prompt. Whilst a functional simulator or training model can be produced rapidly, it can require considerable time to finesse. Model testing is vital. This can first be done by yourself, then also with peers, who may offer a more objective assessment.
It is also necessary to ensure the simulator works on your target device (e.g. phones, computers or both) and this requirement can form part of the prompt. Consider how to provide feedback to the user to aid their learning, or add other useful features that may enhance the experience. For example, with the RAPiD tool we were aware that skin of colour is often underrepresented in literature and education. We were keen to ensure the simulated eyes in RAPiD reflected a diverse patient population, so we added an additional feature to randomly vary the tone of both skin and iris between each case.
Evaluation
Once a viable output is produced that is ready to test with a target audience, it is necessary to consider methods of evaluation, as well as any requirement for ethical approval. We present an initial, simple four-part framework to assess AI co-created learning tools (ACOLT), this can be helpful to hone and summarise what your tool is trying to do before wider testing. We apply this framework to two of our examples: RAPiD and OKNdrum in Table 2.
The best evaluation may be a comparison to existing or traditional methods of learning. A simple preliminary evaluation of the tool by learners can be easily carried out with an anonymous, optional survey. Questions in the survey could include assessment of user-reported self-confidence, proficiency (i.e. using the tool itself to assess whether users are getting better at the task or skill being taught) and experience. For a brief example of preliminary evaluations of learning tools and their findings: wkxm.org/ai-co-creation.
Table 3: ACOLT: A framework for efficient AI co-creation of learning tools, applied to our two examples, a RAPD learning tool and OKNdrum simulator.
Conclusion
We have presented a simple five-step process for the utilisation of AI-based aids to create models, simulators and learning tools, as well as a four-part framework for their evaluation. The process has been demonstrated through two tools we developed, RAPiD and OKNdrum, both of which are freely available online for use by anyone and are potentially suitable in low-resource settings. We noted how ideation is important, but that iteration to refine instruction prompts to achieve the desired output, is key to success. The conclusion is that today’s AI-based aids remove barriers to creating new small educational tools and perhaps leaves our imaginations as the limiting factor. As such AI aids become more capable, we will in turn be able to do more. Is there an unmet need in your day-to-day work? If so, go forth and create!
References
1. Gordon M, Daniel M, Ajiboye A, et al. A scoping review of artificial intelligence in medical education: BEME Guide No. 84. Med Teach 2024;46(4):446–70.
2. Mawyin-Muñoz CE, Salmerón-Escobar FJ, Hidalgo-Acosta JA, Calderon-León MF. Medical simulation: an essential tool for training, diagnosis, and treatment in the 21st century. BMC Med Educ 2025;25(1):1019.
3. Walker M. Teaching in theatre. In: Peyton JWR (Ed.). Teaching and learning in medical practice. Rickmansworth, UK; Manticore Europe Limited; 1998:171–80.
4. Yang C-W, Wang H-C, Chou EH-C, Ma MH-M. Fidelity does not necessarily result in effectiveness – A randomized controlled study in a simulation-based resuscitation training for medical students. Resuscitation 2012;83(1):E116.
5. Massoth C, Röder H, Ohlenburg H, et al. High-fidelity is not superior to low-fidelity simulation but leads to overconfidence in medical students. BMC Med Educ 2019;19(1):29.
Declaration of competing interests: None declared.
