Prospective and External Validation of an Ensemble Learning Approach to Sensitively Detect Intravenous Fluid Contamination in Basic Metabolic Panels

Background Intravenous (IV) fluid contamination within clinical specimens causes an operational burden on the laboratory when detected, and potential patient harm when undetected. Even mild contamination is often sufficient to meaningfully alter results across multiple analytes. A recently reported unsupervised learning approach was more sensitive than routine workflows, but still lacked sensitivity to mild but significant contamination. Here, we leverage ensemble learning to more sensitively detect contaminated results using an approach which is explainable and generalizable across institutions. Methods An ensemble-based machine learning pipeline of general and fluid-specific models was trained on real-world and simulated contamination and internally and externally validated. Benchmarks for performance assessment were derived from in silico simulations, in vitro experiments, and expert review. Fluid-specific regression models estimated contamination severity. SHapley Additive exPlanation (SHAP) values were calculated to explain specimen-level predictions, and algorithmic fairness was evaluated by comparing flag rates across demographic and clinical subgroups. Results The sensitivities, specificities, and Matthews correlation coefficients were 0.858, 0.993, and 0.747 for the internal validation set, and 1.00, 0.980, and 0.387 for the external set. SHAP values provided plausible explanations for dextrose- and ketoacidosis-related hyperglycemia. Flag rates from the pipeline were higher than the current workflow, with improved detection of contamination events expected to exceed allowable limits for measurement error and reference change values. Conclusions An accurate, generalizable, and explainable ensemble-based machine learning pipeline was developed and validated for sensitively detecting IV fluid contamination. Implementing this pipeline would help identify errors that are poorly detected by current clinical workflows and a previously described unsupervised machine learning-based method.