Interpreting antibody identification tests with large language models: Critical role of tabular data formatting

Area of Science:

• Transfusion Medicine
• Computational Biology
• Artificial Intelligence in Healthcare

Background:

• Antibody identification is critical for preventing transfusion reactions.
• Automated interpretation methods struggle with complex antibody identification cases.
• Large language models (LLMs) present a potential solution, but their efficacy with tabular antigen profile data is uncertain.

Purpose of the Study:

• To evaluate the accuracy of various large language models (LLMs) in interpreting simulated antibody identification cases.
• To compare the impact of different data formats on LLM performance for antigen profiles.
• To assess the potential of LLMs as assistive tools in transfusion medicine.

Main Methods:

• Generated 100 simulated antibody identification cases (single/dual alloantibodies, autoantibodies, unidentifiable patterns).
• Evaluated seven LLMs (including GPT-5, Gemini 2.5 Pro) using five text-based data formats and spreadsheet uploads.
• Assessed accuracy across models, formats, and antibody categories through 10 interpretations per case/format.

Main Results:

• Reasoning-capable LLMs achieved a median accuracy of 84.5%, significantly outperforming non-reasoning models.
• Sentence-based serialization of antigen profiles yielded the highest accuracy among text formats; CSV performed worst.
• GPT-5 and o3 showed comparable high accuracy when antigen profiles were uploaded as spreadsheets.

Conclusions:

• Reasoning LLMs possess baseline capability for antibody identification interpretation with appropriate data structuring.
• LLM performance is highly sensitive to data representation and case complexity.
• Current LLM application in transfusion practice should be restricted to an assistive role under expert supervision.