Using Generative AI to Support Patient Information Leaflet Development: A Descriptive Implementation Study

Introduction

Patient information leaflets (PILs) help explain diagnoses, treatments, and ways patients can manage their own health. Creating a resource for patients involves preparing medical information, checking that it is correct and easy to read, designing its appearance, and getting approval through local healthcare procedures. These tasks may involve several groups of professionals, but this study did not measure how long they took or how much they cost.

Natural language processing has become an important part of medical informatics, with uses in healthcare and other settings [1]. AI has also been used for visual design. Lin et al. described AutoPoster, a system that automatically creates advertising posters based on their content [2]. Although this system was not designed for healthcare, it shows another way AI can help with design. Alowais et al. reviewed the use of AI in clinical practice [3]. Generative AI can create text or visual content. Large language models (LLMs) are a type of generative AI designed for language tasks. ChatGPT (OpenAI, San Francisco, CA) is an application based on an LLM.

Earlier studies have explored AI for dental patient education and communication [4]. Sudharshan et al. studied whether ChatGPT could make patient education texts easier to understand [5]. Tukur Jido et al. assessed how easy it was to read AI-generated leaflets about dementia-related conditions [6]. Naufal et al. compared brochures about breast cancer surgery created by ChatGPT and Gemini [7]. A case study has also discussed the use of AI in public health [8]. Together, these publications provide background for studying how AI can help develop patient information.

Lymphoedema is a common, long-term health condition that can have important effects on patients. Clear information about its symptoms, management, and self-care is valuable [9,10]. Wider reviews have discussed both the possible benefits and the concerns surrounding ChatGPT in healthcare education and practice [11]. Studies of ChatGPT's performance in medical examinations show another possible educational use for LLMs. However, doing well in an examination does not prove that a patient leaflet is good quality [12]. The benefits, limits, and risks of GPT-4 in medicine have also been discussed [13]. The World Health Organization has issued guidance stressing the need for ethical oversight and proper governance when AI is used in healthcare [14].

Studies more directly related to patient information include a review by Nasra et al., who brought together research on AI and the readability of patient education materials [15]. Sridharan et al. assessed patient instructions and educational materials created by LLMs, using measures such as how easy they were to understand and how well they helped patients know what actions to take [16]. Sivaramakrishnan et al. compared AI-generated dental education materials by looking at readability, understandability, actionability, and reliability [17]. These studies assess the quality of AI-generated content. However, the practical steps of drafting a leaflet, checking and revising its medical content, designing it, and getting approval through an NHS patient-information governance process have been described less often.

This report aimed to describe how AI was used to develop one patient information leaflet about lymphoedema and how the leaflet moved through the Trust's approval process. Successful implementation was defined as completing the leaflet and receiving formal approval for clinical use.

Technical Report

Study design and implementation outcomes

This was a descriptive report of a proof-of-concept project involving one patient information resource at Imperial College Healthcare NHS Trust. Successful implementation was defined in three ways: (i) completing a PIL with AI assistance, (ii) submitting it through the existing governance process, and (iii) receiving approval for clinical use. These outcomes were described in words. The study did not use a validated measure of feasibility, compare the process with conventional leaflet development, assess costs, or measure patient outcomes.

AI-assisted text generation and clinical review

The free version of ChatGPT (OpenAI) was used to create the first draft of the patient-facing text. The exact model version and date of the original interaction could not be confirmed afterwards. The first prompts asked for plain-language explanations of lymphoedema, its causes, symptoms and signs, diagnosis, management and prevention, and common questions. After reviewing the initial answers, ChatGPT was asked to combine the information into a shorter, easy-to-read leaflet. The exact prompts are included in the Appendix. The prompts had not been formally tested or validated.

The authors reviewed the generated draft before sending it for institutional approval. They checked medical accuracy, completeness, suitable language for patients, and relevance. They made medical corrections, simplified the wording, added useful information, and edited the text. The source documents currently available do not identify the exact medical statements that were corrected. Therefore, this report does not attribute any particular error to ChatGPT without a verified comparison between the first draft and the approved version.

Visual design using Canva

The reviewed text was moved into Canva for visual design. Canva templates, icons, layout tools, and AI-assisted design features helped present the information. Manual changes were also made to the formatting and layout so that the leaflet met the Trust's standards for patient information. ChatGPT was used to draft the text, while Canva supported the visual design process.

Institutional governance and approval

The leaflet was submitted through Imperial College Healthcare NHS Trust's established process for approving patient information materials. The original manuscript states that the review pathway included the Patient Editor, Divisional Quality and Safety Department, Trauma Q&A meeting, and Consultative Expert Working Group on Research and Development (CEWG). The formal governance review did not request any specific changes. This is separate from the medical and editorial changes made before the leaflet was submitted.

Results

The first complete text and visually designed PIL were produced in about one hour. This estimate covers the drafting and design stages only. It excludes the later formal governance process and does not show how much time was saved compared with conventional leaflet development.

The first draft generated by ChatGPT is shown in Figures 1-2. Medical and editorial review led to corrections, simpler terms, additional information for patients, and changes to the layout. These changes show that human involvement went beyond simply placing the text into an institutional template.

The final approved leaflet is shown in Figure 3. After the Trust's governance review, the leaflet was approved for clinical use without any specific additional changes being requested. All three implementation outcomes were achieved for this one resource: completion, submission through the governance process, and approval. No formal assessment of patient usability, readability, understanding, or clinical outcomes was carried out.

Discussion

This report describes how one patient information leaflet about lymphoedema was created and approved using AI-assisted writing and design within an existing NHS governance process. Its main contribution is a clear account of the steps taken, rather than evidence that this method works better or faster than traditional leaflet development.

Previous research provides a reason to explore this approach carefully. Sudharshan et al. studied how ChatGPT could simplify patient education materials [5], while Tukur Jido et al. and Naufal et al. assessed different features of AI-generated leaflets and brochures [6,7]. Nasra et al. reviewed evidence about AI and the readability of patient education materials [15]. Other studies used structured assessment tools to measure how understandable and actionable these materials were [16,17]. These studies focus on the quality of the finished content, rather than the specific governance outcome described in this report.

Human input played an important role in the process. Before the leaflet was submitted for approval, the authors corrected medical information, simplified the language, added relevant content, and changed the presentation. These changes matter because a draft that reads smoothly is not necessarily a medically reliable patient resource. However, the available text does not fully record the exact medical changes between the first draft and the final version. Therefore, this report cannot measure how many AI-generated errors occurred or how serious they were. Future projects should keep a structured record of all changes.

ChatGPT and Canva had different roles. ChatGPT helped produce the first text draft. Canva, including its AI-assisted features, helped with visual formatting and design, which were then adjusted manually.

The reported one hour spent on drafting and design should not be taken as proof of improved efficiency. The study did not measure the time or cost of developing a leaflet using conventional methods, and the governance process took place separately. Similarly, approval by the institution shows that the leaflet passed its local quality-checking process. It does not prove that patients understood the information better, found it easier to access or act on, or experienced better health outcomes.

Using generative AI in healthcare raises questions about accuracy, transparency, responsibility, and professional oversight. Wider clinical research discusses the benefits and risks of language models [11,13], while World Health Organization guidance stresses the need for appropriate governance when AI is used in healthcare [14]. In this project, people remained responsible for reviewing and approving the information given to patients rather than relying on AI-generated content without checking it.

The study has several limitations. It involved only one leaflet at one NHS Trust, had no comparison with conventional methods, included no cost analysis, did not formally validate the prompts, and could not confirm which ChatGPT model had been used. Details of the AI features used in Canva were also incomplete. The study did not formally assess readability, patient usability, or clinical outcomes. Future research should record prompt wording and model versions in advance, measure the time spent at each stage, document clinical changes, and evaluate the materials from the patient's perspective using established assessment tools.

Conclusions

This descriptive report documents the successful development and approval of one patient information leaflet about lymphoedema using ChatGPT-assisted writing and Canva-assisted visual design within an NHS Trust. Human medical and editorial review remained an essential part of the process. The findings show that the workflow was completed successfully in this setting, but they do not prove that it saves time or money compared with conventional methods, can be used successfully in other settings, or improves patient outcomes. Further studies comparing different approaches and evaluating patient experiences are needed.