Multivariate anomaly detection models enhance identification of errors in routine clinical chemistry testing

Area of Science:

• Clinical Chemistry
• Laboratory Medicine
• Data Science in Healthcare

Background:

• Conventional autoverification rules assess analytes individually, risking missed errors from complex patterns like serum contamination.
• Serum contamination by collection tube additives can lead to inaccurate test results.

Purpose of the Study:

• To evaluate the effectiveness of multivariate anomaly detection algorithms in identifying serum contamination and single-analyte errors.
• To compare the performance of multivariate models against traditional limit checks in laboratory autoverification.

Main Methods:

• Developed and compared multivariate Gaussian, k-nearest neighbours (KNN) distance, and one-class support vector machine (SVM) models against conventional limit checks.
• Utilized a dataset of 127,451 electrolyte, urea, and creatinine (EUC) results for training and evaluation.
• Assessed model performance in detecting samples spiked with common collection tube additives (EDTA, fluoride, citrate) and simulated single-analyte errors.

Main Results:

• KNN distance and SVM models significantly outperformed limit checks in detecting all tested contaminants.
• Multivariate Gaussian models showed superiority in detecting most additives, except EDTA.
• All tested multivariate models demonstrated better performance than limit checks in identifying single-analyte errors, with KNN distance showing the highest sensitivity.

Conclusions:

• Multivariate anomaly detection models, especially KNN distance, offer superior error detection capabilities for laboratory autoverification.
• Implementing multivariate approaches in autoverification can optimize error detection, reduce false results, and enhance patient safety.