A Curated Dataset for Question Answering on Laboratory Test Reference Ranges and Interpretation.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Laboratory Science

Background:

• Laboratory test interpretation relies on reference ranges, which vary based on factors like age, gender, and specimen type.
• Automated clinical decision support systems face challenges in accurately interpreting diverse lab test results.
• Accurate interpretation is vital to prevent misdiagnoses and ensure appropriate clinical decisions.

Purpose of the Study:

• To introduce LabQAR, a manually curated dataset of laboratory test reference ranges.
• To evaluate the performance of various large language models (LLMs) in interpreting lab test results.
• To assess the potential of LLMs for automated clinical decision support in laboratory diagnostics.

Main Methods:

• Developed LabQAR, a dataset with 550 reference ranges for 363 unique lab tests, including annotated multiple-choice questions.
• Tested multiple large language models (LLMs), including LLaMA 3.1, GatorTronGPT, GPT-3.5 turbo, GPT-4 turbo, and GPT-4o.
• Evaluated LLM performance in predicting reference ranges and classifying results as normal, low, or high.

Main Results:

• GPT-4o achieved the highest performance among the evaluated LLMs.
• The study demonstrated the capability of LLMs to predict reference ranges and classify lab results.
• LabQAR dataset provides a valuable resource for developing and benchmarking AI models in this domain.

Conclusions:

• Large language models show significant promise for enhancing automated clinical decision support systems.
• GPT-4o exhibits strong potential for accurate laboratory test result interpretation.
• Further development and validation of LLMs are warranted for clinical applications in laboratory medicine.