Detecting Blood Transfusion Adverse Events from Clinical Text Using Domain-Specific BERT

Blood transfusion is a life-saving intervention in critical care that is not without risk, seeing that adverse transfusion reactions can be found in clinical notes. Although systematic fields frequently overlook complex symptoms, narrative information includes fever, hypotension, or hemolysis that are strongly indicative of a TRAE. Here, we use domain-specific BERT models to identify TRAEs from free-text narratives using the MIMIC-III dataset. We retrieve transfusion records and temporally align them with clinical notes to delineate a context window around when the adverse event is most likely to be expressed. Notes are weakly labeled with a curated lexicon of TRAE indicators and fine-tuned ClinicalBERT for binary classification. Our model attains precision of 88.7%, recall of 85.2%, an F1-score of 86.9%, and an AUROC of 93.1% on the held-out test set. Inference latency is low at 58 ms per note, enabling real- or near-real-time integration into hospital workflows. Interpretability analysis reveals that medical terms including "rash," "febrile," and "drop in BP" are driving predictions of the trained model, instilling confidence in the quality of clinical operation. In the future, we believe that this framework can be extended to larger datasets like MIMIC-IV and eICU for facilitating scaling across institutions. In addition, the model can be served via secure API endpoints for a scalable AI-augmented transfusion surveillance infrastructure in a privacy-aware system to enhance clinical vigilance where it matters the most.