Jiwoo Lee, Yousun Chung, Dae‐Hyun Ko, Eun Young Song, Hyungsuk Kim
Reasoning LLMs demonstrated measurable baseline capability for interpreting antibody identification results when accompanied by appropriately structured antigen profile inputs. However, performance varied markedly with data representation and case complexity, and the use of LLMs in transfusion practice should remain limited to an assistive role under specialist supervision.
BACKGROUND: Antibody identification testing is essential for preventing hemolytic transfusion reactions from unexpected red cell antibodies. Although automated interpretation is desirable to improve efficiency and consistency, conventional algorithmic approaches show limited accuracy, particularly in complex cases. Large language models (LLMs) offer a potential alternative; however, their performance is known to be sensitive to tabular data, which is a standard format for antigen profiles in antibody identification. STUDY DESIGN AND METHODS: We generated 100 simulated antibody identification cases comprising single alloantibodies, dual alloantibodies, autoantibodies, and unidentifiable patterns. Seven LLMs (GPT-4o, GPT-4.1, o3, GPT-5, Gemini 2.0 Flash, Gemini 2.5 Flash, and Gemini 2.5 Pro) were evaluated. Antigen profile tables were provided using five text-based formats (Markdown, CSV, JSON, list-based, and sentence-based serialization) and as uploaded spreadsheet files. Each model interpreted all cases 10 times per format, and accuracy was assessed overall and by antibody category. RESULTS: Reasoning-capable LLMs significantly outperformed non-reasoning models, achieving a median accuracy of 84.5%. Among text-based formats, sentence-based serialization yielded the highest accuracy, whereas CSV performed poorest. The maximum accuracy difference between table formats reached 40.5 percentage points. GPT-5 and o3 performed comparably when antigen profiles were supplied as uploaded spreadsheet files, achieving accuracy similar to that of the sentence-based format. CONCLUSIONS: Reasoning LLMs demonstrated measurable baseline capability for interpreting antibody identification results when accompanied by appropriately structured antigen profile inputs. However, performance varied markedly with data representation and case complexity, and the use of LLMs in transfusion practice should remain limited to an assistive role under specialist supervision.