The Clinicians' Guide to Large Language Models: A General Perspective With a Focus on Hallucinations

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Medical Technology Assessment

Background:

• Large language models (LLMs) are advanced AI tools with transformative potential across all medical disciplines.
• Growing interest exists among healthcare stakeholders to integrate LLMs into routine clinical practice.
• Awareness of LLM risks, particularly "hallucinations" (generated false information), is crucial for safe implementation.

Purpose of the Study:

• To highlight the significant risk of hallucinations posed by LLMs in medical applications.
• To explain the multifaceted causes of LLM hallucinations, including training data and model architecture.
• To propose a technical framework for the safe clinical and institutional adoption of LLMs.

Main Methods:

• Analysis of LLM capabilities and limitations in a medical context.
• Identification of potential clinical implications of LLM-generated inaccuracies.
• Development of a general technical framework for LLM risk mitigation.

Main Results:

• LLM hallucinations can lead to inaccurate diagnostic and therapeutic information.
• These inaccuracies may reinforce flawed clinical reasoning and reduce reliability.
• A structured framework is proposed to address and mitigate LLM risks.

Conclusions:

• Clinicians must be aware of the inherent risks, such as hallucinations, associated with LLMs.
• A proactive technical framework is essential for the safe and effective integration of LLMs into clinical practice.
• Mitigating LLM risks is key to realizing their potential benefits in medicine.