An AI-driven process for designing healthcare training scenarios

The cornerstone of effective healthcare simulation lies in meticulously crafted scenarios that provide realistic, controlled environments for learners to develop clinical skills and practice decision-making [1–3]. Traditionally, the creation of these high-fidelity scenarios has been a resource-intensive process, demanding significant time and expertise from simulation educators. The often-quoted figure of 24 h of preparation for a 10–20 min scenario underscores the substantial investment required [1, 4]. The recent advent of advanced large language models (LLMs), such as ChatGPT, has sparked considerable interest in their potential to revolutionize this process [5–8]. Early explorations suggest that LLMs can contribute to streamlining scenario development, potentially offering insights that conventional methods might miss. However, concerns regarding the accuracy, relevance, and structural coherence of AI-generated content have also been raised, necessitating careful consideration and human oversight [9].

This article chronicles the development and implementation of an innovative, AI-powered, agentic workflow for healthcare simulation scenario design in SIMNOVA Simulation Center in Novara, Italy. The project evolved from an initial prototype based on a structured ChatGPT interface to a fully automated system leveraging “n8n,” a no-code open-source platform. The narrative details the development process, the underlying rationale, the scope of potential applications, and the critical lessons learned. The overarching objective is to provide simulation educators with a comprehensive understanding of how AI-driven approaches can enhance efficiency and consistency in scenario design while ensuring rigorous adherence to established best-practice standards.

The primary aim of this innovation was to develop an AI-powered system capable of significantly reducing the time and effort required to create high-quality, standards-compliant healthcare simulation scenarios. This reduction in development time was hypothesized to free up valuable educator resources, allowing for greater focus on learner interaction and debriefing [10].

Beyond simply accelerating the process, the project aimed to achieve several specific objectives. First, the system needed to automate the generation of core scenario components, including learning objectives, a detailed patient narrative, relevant diagnostic data, and comprehensive debriefing points. Second, it was essential that the system’s outputs were consistently aligned with established simulation design standards, specifically the INACSL Standards of Best Practice: Simulation Design and the ASPiH Standards Framework for Simulation-Based Education [11, 12]. Third, user-friendliness was paramount; the workflow needed to be accessible to educators with varying levels of technical expertise, not requiring specialized programming skills. Fourth, the system was designed for adaptability, supporting customization for a wide range of clinical settings and learner levels, from undergraduate nursing students to experienced medical professionals. Fifth, to enhance its global applicability, the system incorporated multilingual scenario generation capabilities. Finally, the development process prioritized addressing potential issues of factual accuracy, bias in AI outputs, and ethical considerations related to intellectual property and patient privacy.

The primary target group for this innovation encompasses a broad spectrum of individuals involved in healthcare simulation. This includes simulation educators across various disciplines (nursing, medicine, allied health professions), simulation center directors and technical staff responsible for scenario implementation, and curriculum developers seeking to integrate simulation-based learning into their programs. Furthermore, the system is particularly relevant for institutions aiming to expand their simulation offerings and for educators working in resource-constrained settings where time and expertise may be limited. The system’s accessibility aims to lower the barrier to entry for those new to simulation scenario design, while also providing advanced capabilities for experienced simulationists.

The agentic workflow within n8n was designed to mimic the structure of a complex, multi-step reasoning process, drawing heavily on principles of agent-based AI systems [16].

The core concepts employed include the following: Decomposition (sub-agents): The overall task of scenario generation was broken down into a series of smaller, more manageable sub-tasks. Each sub-task was handled by a dedicated agent, implemented as a separate node or set of nodes within the n8n workflow. This hierarchical decomposition allowed higher-level agents to orchestrate lower-level, specialized agents, ensuring that each component of the scenario received focused attention. Prompt chaining (sequential execution with context passing): The workflow utilized prompt chaining extensively, where the output of one agent became part of the input prompt for the subsequent agent. This ensured that each agent had the necessary context from previous steps. This serial prompt chaining created a coherent progression of scenario elements. Parallelization (concurrent execution): Where appropriate, agents were executed in parallel to improve efficiency. For example, agents responsible for generating “roles and resources” and “briefing, debriefing, references” could operate concurrently, as they both depended on common inputs but were independent of each other. This parallel execution significantly reduced the overall processing time. Retrieval-augmented generation (RAG): The initial interaction agent utilized RAG to access and incorporate information from various knowledge sources, including a vector database populated with healthcare simulation textbooks, direct search capabilities for PubMed and Semantic Scholar, and guidelines repositories for specific clinical scenarios. This enhanced the factual accuracy and relevance of the generated content. Tool use (external API calls and data transformations): The workflow leveraged n8n’s ability to integrate with external APIs to perform tasks such as retrieving relevant research articles, generating text with LLMs, validating data formats, converting data between formats, and sending emails. This integration capability expanded the system’s functionality beyond what a standalone LLM could achieve. Iterative refinement (reviewer agent and feedback loop): A dedicated “reviewer” agent was included in the workflow to assess the output of the scenario generation process for adherence to the user’s initial query and established simulation best-practice standards. This provided a crucial layer of quality control and allowed for potential automated refinement before the final output was presented to the user.

In parallel with the main workflow, supplemental material creation agents generated realistic clinical documents (e.g., arterial blood gas results, imaging findings, laboratory test results) based on the patient’s condition and progression defined in the scenario states. Finally, an assembly process combined the outputs from various agents into a single, cohesive document, which was then delivered to the user’s email address.

The entire workflow represented a sophisticated orchestration of AI agents, each with specialized responsibilities yet working in concert to produce a comprehensive, standards-compliant simulation scenario. This approach not only accelerated the scenario creation process but also enhanced consistency and adherence to established guidelines.

The AI-driven scenario design workflow possesses a broad scope of potential application within the field of healthcare simulation. The system’s versatility allows it to be adapted to a wide range of clinical settings, including, but not limited to, emergency medicine, critical care, obstetrics, pediatrics, and mental health. Furthermore, it can be customized to suit different learner levels, ranging from undergraduate students to experienced healthcare professionals, by adjusting the complexity of the clinical presentation and the expected level of performance.

The system’s design facilitates the tailoring of scenarios to specific learning objectives, encompassing clinical skills acquisition, teamwork and communication training, and clinical decision-making practice. The inclusion of multilingual output capabilities significantly expands the system’s reach, making it applicable in diverse educational contexts across the globe. The workflow also facilitates the rapid generation of scenario variations, allowing educators to create multiple versions of a scenario with differing patient presentations, complications, or resource availability. This capability is particularly valuable for research purposes, enabling controlled comparisons of different pedagogical approaches. The generated scenarios can be used in simulation with standardized patients, and within environments such as hospital wards, simulation centers, and community settings.

Implementing and utilizing this innovation requires a series of straightforward steps. First, access to the n8n platform is necessary, either through a self-hosted instance or a cloud-based subscription. The pre-built scenario design workflow can then be imported into the n8n environment. API keys for the required AI models, such as OpenAI’s GPT models, Anthropic’s Claude, and Google’s Gemini models, must be obtained and configured within the n8n workflow.

To generate a scenario, educators provide a clear and concise brief, specifying the clinical topic, target learner level, learning objectives, and any specific requirements or constraints. The workflow is then executed within n8n, triggering the sequence of AI agent actions. Critically, after the workflow completes, a thorough review and refinement of the AI-generated scenario by a subject matter expert is essential. This step ensures the accuracy, realism, and pedagogical soundness of the scenario before it is used with learners. Finally, the finalized scenario can be implemented in simulation-based education activities.

For simulation centers with sufficient computational resources, the system can be configured to use locally-hosted, smaller LLMs rather than commercial API services. This option reduces ongoing operational costs and addresses potential data privacy concerns, making the system more accessible to resource-constrained institutions. The implementation typically requires modest technical skills comparable to those of a simulation technician or educator with basic coding experience.

While the AI-driven workflow offers significant advantages, several potential pitfalls must be acknowledged and addressed. A primary concern is the potential for inaccuracies in the AI-generated content. LLMs, while powerful, do not possess true understanding and may occasionally produce information that is factually incorrect or clinically inappropriate. To mitigate this risk, a rigorous review process involving subject matter experts is crucial.

Another potential issue is bias in the AI model’s outputs, reflecting biases present in the training data. To address this, prompts were designed to encourage diverse and equitable representation, and the outputs are reviewed for inclusivity. Technical issues with the n8n platform or the AI models themselves may also arise, necessitating backup plans and access to technical support. For centers using self-hosted LLMs, additional challenges may include managing computational resources and maintaining model updates. These can be addressed through careful infrastructure planning and establishing regular update protocols. The initial learning curve for configuring n8n workflows might also present a barrier, though this can be mitigated by providing detailed documentation and establishing peer support networks among simulation centers implementing similar systems.

The resources required for implementation include access to the n8n platform, which may involve costs depending on the chosen deployment option (self-hosted or cloud-based). API access to the chosen AI models also incurs costs, which are dependent on usage and the specific provider’s pricing structure. For centers opting for self-hosted n8n and local LLMs, initial costs would include server hardware with sufficient computational capacity, particularly if running multiple concurrent workflows.

A precise cost estimation is challenging, as it is highly dependent on usage patterns and specific pricing plans. However, it is important to consider the potential cost savings resulting from the significant reduction in scenario development time, which may offset the direct costs of the platform and API access. By our preliminary estimations, the system reduces scenario development time by approximately 70–80%, translating to substantial labor cost savings for simulation centers regularly producing new scenarios. This aligns with the findings by the University of Toronto’s pilot program, which showed reduced scenario development time by 73% while maintaining 92% clinical accuracy compared to human-authored cases when using AI as an aid to writing [18].

The self-hosted approach with local LLMs offers the most cost-effective long-term solution for centers with frequent scenario development needs, as it eliminates ongoing API fees. However, this approach requires greater initial investment in hardware and technical configuration. For centers with occasional scenario development needs, the cloud-based approach with commercial API services may offer a more flexible, pay-as-you-go solution.

The strategic use of different AI models in our implementation demonstrates how costs can be managed effectively. By utilizing Gemini 2.0 models (which are currently free to use) for the main workflow components while reserving premium models like GPT-4o and Claude 3.5 for specialized tasks requiring their unique capabilities, the system achieves cost optimization without compromising quality.

The development and implementation of this AI-assisted scenario design workflow yielded several valuable lessons. The importance of interdisciplinary collaboration between simulation experts and individuals with technical aptitude was paramount, though we demonstrated that specialized AI expertise is not a prerequisite for successful implementation. The iterative process of refining AI prompts and workflow logic proved crucial for optimizing the quality and accuracy of the outputs. The project reinforced the essential role of human oversight in AI-assisted processes, emphasizing that AI should be viewed as a powerful tool, not a replacement for human expertise. The need for user training became evident, as educators require guidance on how to effectively interact with the AI tool and interpret its outputs. Ethical considerations, particularly regarding accuracy, bias, and intellectual property, required careful attention throughout the development process. Finally, the process of creating the AI workflow unexpectedly led to a deeper understanding of our own scenario design practices, forcing us to make our implicit knowledge explicit and transferable.

Perhaps most significantly, the project demonstrated that sophisticated AI implementations in healthcare simulation are within reach of motivated clinicians and educators without requiring specialized technical expertise or external consultants. This realization has profound implications for the democratization of AI tools within healthcare education, potentially accelerating innovation and adoption across diverse settings and resource levels.