Carly Maucione, MD, a resident physician at Washington University in St. Louis

Carly Maucione, MD

Uncovering Hidden IV Fluid Contamination in CBCs, With Carly Maucione, MD

Uncovering Hidden IV Fluid Contamination Through Machine Learning, With Carly Maucione, MD

Machine learning may help close the gap in identifying intravenous (IV) fluid contamination in complete blood counts (CBCs) and prevent unnecessary transfusions, according to new research.

The findings from the multicenter study suggest IV fluid contamination in CBCs may be more common than previously recognized, with machine learning models identifying possible contamination in approximately 2% of samples.

“Finding that there is this unmet need, I think, should shock some people, and I think it’s okay to be shocked,” said study investigator Carly Maucione, MD, a resident physician at Washington University in St. Louis, in an interview with 

. “The results were higher than we expected. But there’s mounting evidence that IV fluid contamination is a significant issue, and it’s something we could address in the laboratory.”

Specimen contamination with IV fluids has long been recognized as a challenge in clinical laboratories, as it can produce erroneous CBC results, influencing diagnostic interpretation and clinical decision-making.

As Maucione explained, in inpatient settings, clinicians may encounter unexplained drops in hemoglobin that later normalize, sometimes without a meaningful impact on patient management. In more concerning scenarios, however, falsely low hemoglobin values may prompt unnecessary transfusion decisions.

Still, identifying IV fluid contamination after samples arrive in the laboratory remains difficult. Detection requires laboratory technologists' expertise, with a lack of standardized methods to train pathologists. As such, recognition depends largely on a technologist’s ability to identify abnormal result patterns and infer possible causes. This approach is inherently variable and difficult to scale across high-volume laboratory workflows.

Study investigators recognized machine learning as a potential solution to standardize care and identify IV fluid contamination with precision and accuracy.

To address this gap, investigators conducted a retrospective, multicenter machine learning diagnostic validation study using real-world inpatient data from 2 institutions. The team developed and tested 2 machine learning models designed to retrospectively identify IV fluid contamination in CBC results. Since no gold standard exists for confirming contamination, model outputs were validated against expert chart review.

The models were trained using simulated IV fluid contamination scenarios. They incorporated prior, current, and post-hemoglobin concentrations, platelet counts, and white blood cell counts. Performance was then assessed using 1 year of inpatient CBC data from each institution to evaluate real-world applicability. Transfusions were only classified as potentially unnecessary when post-transfusion hemoglobin values were unexpectedly higher than pre-transfusion values and exceeded 8 g/dL, aligning with commonly used transfusion thresholds.

The models demonstrated strong discriminatory performance, 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, respectively. Across both institutions, approximately 2% of inpatient CBC trios were classified as potentially contaminated, as previously mentioned.

Importantly, the investigators evaluated the clinical implications associated with contaminated results. Among inpatient transfusions for which a CBC trio was available, investigators deemed 6%-9% potentially unnecessary based on a conservative, rule-based definition validated through expert chart review.

“We are seeing that there are clinical consequences, and those clinical consequences are things that we can track using an algorithm such as this one,” said Maucione. “Both of those are things that we think are worth addressing and moving forward in either developing similar algorithms to detect this prospectively, using these as measures in order to track these rates.”

Maucione C, McLamb N, Zaydman MA, Pearson LN, Metcalf RA, Spies NC. Identification of IV fluid contamination in complete blood counts and subsequent unnecessary red blood cell transfusions using artificial intelligence. 

. Published online August 2026:10.1111/trf.70072. doi:https://doi.org/10.1111/trf.70072

Newbigging A, Landry N, Brun M, et al. New solutions to old problems: A practical approach to identify samples with intravenous fluid contamination in clinical laboratories. 

. 2024;127-128:110763. doi:https://doi.org/10.1016/j.clinbiochem.2024.110763

Videos

Machine models identified 2% of CBCs as possibly contaminated and flagged 6–9% of inpatient transfusions as potentially unnecessary, explains Maucione. 


Carly Maucione, MD

Articles