Classification of posttransfusion adverse events using a publicly available artificial intelligence system.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Transfusion Medicine

Background:

• Accurate classification of transfusion reactions is crucial for patient care, donor management, and hemovigilance.
• Generative artificial intelligence (AI) offers potential for diagnostic assistance in healthcare settings.

Purpose of the Study:

• To compare the diagnostic accuracy of a generative AI system against transfusion medicine (TM) specialists in classifying transfusion reactions.
• To evaluate AI's ability to diagnose, assess severity, and determine imputability of transfusion reactions using established criteria.

Main Methods:

• An AI system evaluated 36 transfusion reaction case scenarios.
• AI responses were compared to an expert panel's classifications and prior TM specialist assessments.
• AI performance was benchmarked against TM specialists using the TrDDx web-based algorithm for challenging cases.

Main Results:

• AI accuracy varied significantly by transfusion reaction category, excelling in some (e.g., TACO, TRALI) but failing in others (e.g., DSTR).
• Overall AI accuracy was 48.7% compared to 72.1% for TM specialists (p=0.005).
• AI responses often used non-standard terminology and lacked severity and imputability assessments.

Conclusions:

• Generative AI may assist clinicians in identifying missed transfusion reaction categories.
• Current AI output requires careful consideration and validation before clinical application for transfusion reaction classification.