Prospective and External Validation of an Ensemble Learning Approach to Sensitively Detect Intravenous Fluid...

Area of Science:

• Clinical Chemistry
• Machine Learning in Healthcare
• Laboratory Medicine

Background:

• Intravenous (IV) fluid contamination in clinical specimens poses risks to patient safety and laboratory operations.
• Existing detection methods, including unsupervised learning, lack sensitivity for mild but significant contamination.
• There is a need for more sensitive, explainable, and generalizable methods to detect IV fluid contamination.

Purpose of the Study:

• To develop and validate an ensemble-based machine learning pipeline for sensitive detection of IV fluid contamination.
• To improve upon the sensitivity of current clinical workflows and prior machine learning approaches.
• To ensure the developed pipeline is explainable and generalizable across different institutions.

Main Methods:

• An ensemble machine learning pipeline combining general and fluid-specific models was trained and validated.
• Real-world and simulated contamination data were used for training and validation.
• Performance was assessed using in silico simulations, in vitro experiments, and expert review; SHapley Additive exPlanations (SHAP) were used for interpretability.

Main Results:

• The pipeline demonstrated high sensitivity and specificity in both internal (0.858, 0.993) and external (1.00, 0.980) validation sets.
• SHAP values provided clear explanations for contamination-related analytical alterations.
• The pipeline identified more contamination events exceeding error limits compared to current workflows.

Conclusions:

• An accurate, generalizable, and explainable ensemble machine learning pipeline for IV fluid contamination detection was successfully developed and validated.
• This pipeline offers improved sensitivity for detecting subtle contamination events missed by current methods.
• Implementation of this pipeline can enhance laboratory error detection and patient safety.