Automating the Detection of IV Fluid Contamination Using Unsupervised Machine Learning.

Area of Science:

• Clinical Chemistry
• Machine Learning in Healthcare
• Laboratory Medicine

Background:

• Intravenous (IV) fluid contamination is a frequent source of preanalytical error, potentially leading to incorrect treatment and patient harm.
• Current detection methods, such as delta checks and manual review, are inefficient and prone to human error.
• Supervised machine learning for contamination detection is limited by the need for expert-labeled training data.

Purpose of the Study:

• To develop and evaluate an automated, accurate, and practical method for detecting IV fluid contamination in basic metabolic panel (BMP) results.
• To implement a machine learning model that does not require expert-labeled training data.

Main Methods:

• A Uniform Manifold Approximation and Projection (UMAP) model was trained and tested on over 25 million BMP results from 312,000 patients.
• The model utilized a combination of real patient data and simulated IV fluid contamination.
• An "enrichment score" was derived for classification, and UMAP predictions were compared to the existing clinical workflow via expert chart review.

Main Results:

• UMAP embeddings effectively identified outliers indicative of IV fluid contamination.
• The model achieved a positive predictive value (PPV) of 0.78 at a flag rate of 3 per 1000 results.
• 58 previously undetected contamination cases were identified, including 56 critical results across 49 BMPs.

Conclusions:

• Automated and accurate detection of IV fluid contamination in BMP results is feasible.
• This approach offers a practical solution without the need for expert-labeled training data.
• The developed method enhances patient safety by identifying critical errors in laboratory testing.