LabQAR: A Manually Curated Dataset for Question Answering on Laboratory Test Reference Ranges and Interpretation

Area of Science:

• Medical Informatics
• Clinical Pathology
• Artificial Intelligence in Healthcare

Background:

• Laboratory tests are vital for medical diagnosis and management.
• Interpreting lab results is complex due to factors like age, gender, and specimen type.
• Automated systems require nuanced data to prevent diagnostic errors.

Purpose of the Study:

• To introduce LabQAR, a curated dataset for laboratory test interpretation.
• To evaluate the performance of large language models (LLMs) in understanding lab test reference ranges.
• To explore the potential of LLMs in clinical decision support systems.

Main Methods:

• Developed LabQAR, a dataset with 550 reference ranges for 363 lab tests.
• Included multiple-choice questions with annotations on influencing factors.
• Assessed LLMs (LLaMA 3.1, GatorTronGPT, GPT-3.5, GPT-4, GPT-4o) on prediction and classification tasks.

Main Results:

• GPT-4o achieved the highest performance among evaluated LLMs.
• Models demonstrated varying capabilities in predicting reference ranges.
• Classification of results (normal, low, high) was also assessed.

Conclusions:

• LLMs, particularly GPT-4o, show significant potential for automating lab result interpretation.
• The LabQAR dataset provides a valuable resource for developing and testing such systems.
• Further development of LLMs can enhance clinical decision support and reduce diagnostic errors.