1 Machine Learning Identifies Unrecognized IV Fluid Contamination of Complete Blood Counts That Motivates Potentially Unnecessary Red Blood Cell Transfusions

Intravenous (IV) fluid contamination is uncommon, but often difficult to detect. When undetected, misinformed clinical decisions may lead to downstream consequences, such as unnecessary blood transfusion, whose frequencies may vary by institution. Machine learning-aided contamination detection in complete blood counts (CBC) can help evaluate this by identifying unnecessary transfusions motivated by erroneous results. Methods A LightGBM model was trained to detect simulated ­contamination using prior-current-post trios of hemoglobin, hematocrit, platelet, and white blood cell counts. At Washington University in St. Louis (WashU), 904,205 CBC trios from prior to 2024 were used to train, and 267,565 CBC trios in 2024 were held-out to evaluate. At the University of Utah, 278,447 trios were used to train, with 110,299 held-out to evaluate. Contamination was simulated in training sets by linear dilution of feature input in the current CBCs. Red blood cell transfusions were joined to CBC trios if their administration time fell between collection of the current and post CBC. Potentially unnecessary transfusions were defined as those in which: 1) the current CBC was predicted as contaminated, 2) the post hemoglobin was >8 g/dL, and 3) the post hemoglobin was higher than the prior. Results The models identified simulated contamination in five-fold cross-validation with an area under the precision-recall curve of 0.99 at both institutions. In test sets, 7,726 (2.9%) and 3,033 (2.7%) were predicted as contaminated at WashU and Utah, respectively. Feature importances by SHAP values were greater for hemoglobin than hematocrit and platelet counts, while white blood cell counts had negligible impact. All time points contributed to model predictions, with current results being slightly more impactful than prior and post. Contaminated trios demonstrated a marked “anomaly-with-resolution” pattern absent from uncontaminated specimens, with significant decreases in the current CBCs that tended to return toward their prior baseline upon follow-up testing. 7.1% of 23,684 inpatient transfusions with a matched CBC trio at WashU and 8.8% of 7,632 inpatient transfusions with a matched CBC trio at Utah were potentially unnecessary. The most common indication for unnecessary transfusions was “hemoglobin <7 g/dL” at both institutions. Conclusions Almost 3% of CBC trios are likely contaminated by IV fluids, substantially higher than prior estimates. Of transfusions for which a CBC was available for prediction, about 8% are likely unnecessary, and motivated by spuriously low, diluted hemoglobin concentrations. A more robust prospective validation that can assess performance across different patient subgroups is necessary. Future work should focus on distinguishing bleeding-related patterns from those specific to contamination in real time, allowing for the implementation of machine learning-based approaches to reduce blood product waste and unnecessary transfusion-associated morbidity.