人工知能を用いた全血球計算における静脈内輸液混入の特定とそれに続く不必要な赤血球輸血

Specimens contaminated with intravenous (IV) fluids can lead to considerable measurement errors in complete blood counts (CBCs), posing challenges for laboratory operations and clinical decision-making. There is no gold-standard method for retrospectively identifying these events, making it difficult to target quality improvement initiatives or optimize laboratory detection protocols.

This study aimed to develop and validate machine learning (ML) models to retrospectively identify IV fluid contamination in CBC results at scale across two institutions. The models were trained on simulated contamination in CBCs using prior, current, and post hemoglobin concentrations, platelet counts, and white blood cell counts, then validated against expert-reviewed datasets. Real-world applicability was assessed using 1 year's worth of CBC results from each institution.

The models effectively discriminated contaminated from non-contaminated results on real-world datasets, with areas under the receiver operating characteristic curve of 0.972 and 0.957, and areas under the precision-recall curve of 0.723 and 0.619. In 1 year of CBC data, ~2% of inpatient CBC trios were predicted as contaminated across both institutions. 6%-9% of inpatient transfusions for which a CBC trio was available were deemed potentially unnecessary using a rule set validated by expert chart review.

The findings support the feasibility of using ML to identify IV fluid contamination in CBC results efficiently and effectively. Further work, including prospective real-world evaluations, targeted quality improvement initiatives, and development of real-time detection models, is necessary before realizing the potential benefits to patient safety, laboratory operations, and patient blood management.