Machine learning for the identification of mild to moderate EDTA contamination of serum samples

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background K2EDTA contamination of serum samples is a preanalytical error that poses a risk to patient safety. Contamination is most frequently mild to moderate, which standard detection procedures often miss. We examined whether machine learning models could improve the detection of K2EDTA contamination. Methods Artificial neural network, decision tree (both simple and complex), extreme gradient boosting, k-nearest neighbours, logistic regression, naïve Bayes, random forest, and support vector machine models were developed. Models were trained using extracted patient results for electrolytes, urea, creatinine, albumin-adjusted calcium, magnesium, and phosphate, with K2EDTA contamination errors simulated in silico. Model performance was evaluated on 300 real-world samples, half of which were intentionally contaminated with mild to moderate amounts of K2EDTA. Model performance was compared with that of limit checks, multi-analyte rules and two novel parameters, the potassium/calcium ratio and the potassium/magnesium ratio. Results All nine machine learning models identified K2EDTA contamination more accurately than standard approaches (p-values