Humans and Large Language Models in Clinical Decision Support: A Study with Medical Calculators.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Medical Informatics

Background:

• Large language models (LLMs) are increasingly evaluated for medical knowledge but their utility in clinical decision-making, specifically for tool selection, is unclear.
• Assessing LLMs' capability to recommend appropriate medical calculators is crucial for safe and effective clinical practice.

Purpose of the Study:

• To evaluate the performance of various large language models (LLMs) in recommending medical calculators compared to human performance.
• To identify the types of errors LLMs make when selecting clinical calculators.

Main Methods:

• Nine LLMs (open-source, proprietary, domain-specific) were tested on 1,009 multiple-choice questions covering 35 clinical calculators.
• LLM performance was compared to human annotators on a subset of 100 questions.
• Error analysis was conducted on the highest-performing LLMs.

Main Results:

• The top-performing LLM achieved 66.0% accuracy in answering questions about medical calculators.
• Human annotators achieved an average accuracy of 79.5%, outperforming the LLMs.
• LLM errors were primarily due to comprehension (49.3%) and calculator knowledge deficits (7.1%).

Conclusions:

• Current large language models (LLMs) do not surpass human performance in recommending medical calculators for clinical decision-making.
• LLMs require significant improvement in comprehension and medical knowledge to be reliable tools for clinical calculator selection.