From paper to pocket: artificial intelligence and local cardiology guideline implementation

Abstract

We read with interest the article by Lyell and Magrabi in the Internal Medicine Journal asking whether it is time to ‘hang up the stethoscope’ in the era of artificial intelligence (AI).1 The authors argue convincingly that discussion should move away from the generalised promise of disembodied AI and instead focus on the specific tools, tasks and contexts in which machine learning (ML) may support clinical practice. We agree with this position and wish to extend it by describing a concrete, narrowly defined application: the use of large language models (LLMs) to provide real-time, point-of-care access to locally approved cardiology guidelines. This use case directly addresses the questions posed by Lyell and Magrabi by defining what the tool does, how it is used and the limited clinical task it supports. Barriers include insufficient time to locate and interpret lengthy documents, limited opportunities for formal education and uncertainty about applying recommendations in real-world clinical contexts.2 AI, specifically LLMs using retrieval-augmented generation (RAG), offers a task-specific mechanism to address these barriers by allowing clinicians to query locally approved guidelines using a defined patient context and receive an evidence-based response at the point of care. Importantly, this application aligns with the framework proposed by Lyell and Magrabi; it represents narrowly scoped automation embedded within existing clinical workflows, rather than an attempt to replace clinician judgment. Such systems function to improve access and interpretation of guidance, not to provide autonomous clinical decisions. In this context, LLMs function to support information retrieval and interpretation, without performing diagnosis or prognostication or making treatment decisions. Hospitals with electronic medical records (EMRs) have improved access to local guidelines using intranet systems.3 However, access alone does not reduce the cognitive burden of interpreting and applying the recommendations. LLMs can synthesise locally approved guidelines and return scenario-specific recommendations at the point of care. Embedding requirements for such tools within hospital systems has been described.4 Fig. 1 illustrates a proposed workflow in which a LLM interfaces within an EMR system. When constrained to approved, structured sources, guideline-specific LLMs have shown high accuracy in controlled evaluations.5 Early deployment may occur via existing guideline applications. LLMs require robust government, validation and testing to ensure patient safety, accuracy and confidentiality. Local versus cloud hosting must be considered to address data sovereignty, cybersecurity and patient confidentiality.6 Institutional policy and regulation are needed to govern data handling and clinical use, including in remote settings. Importantly, Lyell and Magrabi emphasise that clinician trust and acceptance are essential, with LLMs positioned to support and enhance clinical practice rather than replace existing roles and specialties.1 Appropriately governed use of LLMs within hospital systems may support timely, evidence-based cardiovascular care at the bedside, particularly where access to on-site specialist support is limited. In this role LLMs move guidelines from ‘paper to pocket’, allowing for more practical and everyday usage, supporting the widespread implementation of best medical practice. Despite this, further evaluation is required to determine LLM effectiveness, safety and governance requirements. Data sharing not applicable to this article as no datasets were generated or analysed during the current study.

Ähnliche Arbeiten