#89 The use of artificial intelligence to generate discharge correspondence from the emergency department (the AIDED study)

Abstract

<h3>Introduction</h3> Timely and accurate discharge communication is vital for patient safety, yet Emergency Department (ED) discharge summaries are often deprioritised due to clinical pressures. Artificial Intelligence (AI) tools offer a potential solution to streamline this process, but their feasibility and quality in ED workflows remain understudied. <h3>Methods</h3> In this prospective observational study, 13 ED healthcare providers used a purpose-built AI assistant (MedWrite™) to generate 132 anonymised discharge summaries. Time-to-completion, frequency and types of errors, and number of edits were recorded. Each summary was independently reviewed by a senior Emergency Medicine consultant and an experienced GP using a structured tool assessing accuracy, completeness, clarity, and clinical acceptability. Interrater agreement and subgroup analyses were performed. <h3>Results</h3> The median time to generate and approve a summary was 106 seconds (IQR 70–128). Errors were identified in 38% of letters (median 1 error per letter, mostly grammatical or formatting). Reviewer 1 deemed 90.2% of summaries clinically acceptable; Reviewer 2 deemed 76.5% acceptable. Letters were rated as better than typical historical letters in 90.2% (Reviewer 1) and 66.7% (Reviewer 2) of cases. However, interrater agreement on key quality domains was low, particularly regarding completeness and clarity of medication or disposition plans. <h3>Conclusion</h3> AI-generated discharge summaries were produced rapidly and rated as clinically acceptable in the majority of cases, showing potential to reduce documentation burden in the ED. However, inter-reviewer variability underscores the need for standardisation and oversight. Safe integration of AI tools into clinical workflows should maintain a human-in-the-loop approach with clear governance around quality benchmarks and discharge communication standards. <i>*presenting author</i>

Ähnliche Arbeiten