Large language model–based patient handoff tools relevant to intensive care: a scoping review

Abstract

Provider-to-provider patient handoffs are a routine yet complex component of intensive care unit (ICU) workflows and are essential for patient safety. Emerging generative artificial intelligence (AI), particularly large language models (LLMs), may improve clinical communication through automated summarisation, documentation and handoff-related workflows. This scoping review followed PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines and searched six databases (PubMed, IEEE Xplore, Google Scholar, ACM Digital Library, arXiv and medRxiv) on 7 October 2025 for studies published between 14 August 1999 and 6 June 2025. Eligible publications examined LLMs or related language technologies in clinical summarisation, documentation, communication workflows, clinical decision support or patient handoffs. Forty-six studies met the inclusion criteria. Most evaluated LLMs for clinical summarisation, information extraction or decision support, whereas studies directly evaluating LLM-generated patient handoffs were rare. Only one study examined emergency department handoffs and no studies evaluated ICU-specific handoffs. Across the literature, LLMs demonstrated promising performance for clinical summarisation but recurrent challenges included hallucinations, clinically relevant omissions, limited model transparency and reliance on linguistic evaluation metrics rather than patient-centred outcomes. Human expert review was frequently incorporated to assess clinical accuracy and safety. Overall, current evidence suggests that LLMs have considerable potential to support future ICU handoff workflows but substantial evidence gaps remain. Prospective evaluation in real-world ICU settings, with clinically meaningful safety metrics and structured communication frameworks, is needed before widespread implementation.