A-194 Intravenous Fluid Contamination of Basic Metabolic Panels Is Underrecognized by Manual Workflows

Abstract Background Erroneous laboratory results contribute to delays in diagnosis, incorrect treatment decisions, and increased healthcare costs. Of these errors, most occur in the preanalytical phase, before the sample makes it to the laboratory. One particularly common source of pre-analytical error is contamination by intravenous crystalloids (IV fluids). Current workflows for detecting IV fluid contamination rely on manual intervention by technologists, which is labor-intensive, variable between technologists, and may lack sensitivity. Methods 9 627 320 chemistry results were aggregated. IV fluid contamination was simulated by mixing a random sample of results in silico with normal saline (NS), lactated Ringer’s (LR), and their 5% dextrose-containing counterparts (D5NS and D5LR). This simulation procedure was validated by in vitro experiments. Machine learning models were trained to predict contamination and mixture percentage using the simulated results and the prior unaffected result from that patient. A random sample of 100 predicted positives were manually chart reviewed and adjudicated. Predictions from the models on retrospective data were compared to real-world technologist flags. Results The ML pipeline identified 3023, 1488, 161, and 236 results as being likely contaminated with NS, LR, D5NS, and D5LR respectively (Fig. 1). Of these, 684 (23%), 84 (6%), 103 (64%), and 148 (63%) had already been flagged by technologists as contaminated. 100/100 manually-reviewed predictions were confirmed. The median contamination ratio for each fluid detected by the pipeline was lower than the current workflow for NS (95% CI of difference: −0.1 to −0.07) and LR (CI: −0.13 to −0.1). The number-needed-to-test to detect one additional contaminated specimen was 125 samples. Conclusion A machine learning pipeline uncovers a substantial number of potentially contaminated results that were previously undetected using manual workflows, especially among non-D5 fluids.