Tailoring Discharge Summaries to Health Care Providers’ Needs (Part 1 of the Framework and Implementation of AI Tools Project): User-Centered Design Approach

TL;DR

This study creates a framework for designing prompts that help AI generate discharge summaries tailored to healthcare providers' needs, using input from medical professionals.

Contribution

A novel, human-centered framework for creating individualized prompts to improve AI-generated clinical summaries.

Findings

Structure and follow-up were the most emphasized categories in the workshop and questionnaire.

The CO-STAR framework improved prompt clarity and alignment with clinical expectations.

Communication was identified as a new, universally valued category for discharge summaries.

Abstract

Medical discharge letters are critical for continuity of care but often lack clarity and personalization, making it difficult for health care providers to retrieve essential information. While large language models (LLMs) offer potential for automating summary generation, their effectiveness depends heavily on the quality and contextual relevance of the prompts used.

The objective of this study was to develop and describe a human-centered, replicable framework for creating individualized prompts that guide LLMs in generating summaries tailored to the specific needs of health care providers.

A multidisciplinary workshop was conducted at Ghent University Hospital with 26 health care providers from 5 institutions, including hospitals and general practitioner networks. Participants brainstormed ideal summary formats, generating 170 ideas categorized into themes such as “structure,” “medical history,” “medication,” and “follow-up.” These insights informed the development of a 110-item structured questionnaire distributed to 33 participants. Responses were used to generate personalized and generic prompts, refined using the context, objective, style, tone, audience, and response (CO-STAR) framework.

“Structure and layout” (40/170, 23.5%) and “follow-up” (27/170, 15.9%) were the most emphasized categories in the workshop. The questionnaire confirmed the importance of the “follow-up” and “medical history” sections. Prompts were generated per participant and by health care type, incorporating frequently selected responses. The CO-STAR framework was applied to improve prompt clarity and alignment with clinical expectations. Communication emerged as a new category during the workshop and was universally valued in the questionnaire.

This study presents a novel, systematic approach to prompt engineering in clinical artificial intelligence applications. By translating qualitative input into structured, individualized prompts, the framework is designed to improve the usability and relevance of artificial intelligence–generated summaries. It proposes a scalable approach for integrating human-centered design into LLM deployment in health care, with the aim of supporting more accurate, context-aware clinical documentation. However, these outcomes remain to be empirically validated; this study is limited to the design and implementation of a human-centered prompt construction pipeline in a specific multicenter setting.

Figures