Simulation-Based Machine Learning as a Benchmarking Tool for the Retrospective Identification of IV Fluid...

Current laboratory workflows fail to reliably capture intravenous fluid contamination of chemistry results, leading to diagnostic uncertainty and misinformed clinical decisions. These events cause predictable error patterns, but no gold-standard detection method or definition exists, hindering optimization efforts. We sought a scalable detection solution for these errors to guide quality initiatives and research efforts.

We developed an ensemble of LightGBM classifiers to retrospectively identify the typical anomaly-with-resolution pattern seen in contaminated results by simulating contamination from common fluid compositions into authentic patient results. Models were evaluated by cross-validation and manual review, applied to a held-out 1-year validation cohort, and used to benchmark existing delta-check and published detection approaches.

Models produced excellent discrimination between real results and simulated contamination. When applied to real-world data, results flagged as contamination consistently exhibited the expected anomaly-with-resolution pattern and were enriched for abnormal and critical results that often exceeded reference change values across multiple analytes. Predicted contamination rates varied substantially by clinical setting, with the highest rates in emergency departments, operating rooms, and intensive care units. The retrospective machine learning ensemble provides a scalable benchmarking tool for evaluating routine laboratory workflows and published rule-based or delta check methods.

Simulation-based machine learning provides a scalable, high-fidelity "silver standard" for retrospective identification of intravenous fluid contamination. This approach enables robust benchmarking of detection strategies, supports optimization of practical rule-based workflows, and offers actionable insights for laboratory quality improvement and patient safety initiatives.