Analyte Importance Analysis in Machine Learning-Based Detection of Wrong-Blood-in-Tube Errors Using Complete Blood...

Area of Science:

• Laboratory Medicine
• Clinical Pathology
• Medical Informatics

Background:

• Wrong blood in tube (WBIT) is a significant pre-analytical error in laboratory medicine.
• Current detection methods have limitations, leading to undetected mislabeling errors.
• WBIT compromises patient safety and diagnostic accuracy.

Purpose of the Study:

• To evaluate the effectiveness of Random Forest models for detecting WBIT errors.
• To identify the most important complete blood count (CBC) analytes for WBIT detection.
• To develop a practical and interpretable machine learning-based solution for WBIT identification.

Main Methods:

• Analysis of 799,721 patient samples from a German tertiary care center.
• Development of Random Forest models using per-analyte differences between paired samples.
• Evaluation of models using F1 score, AUC, sensitivity, and PPV across various CBC analyte subsets.
• Assessment of analyte importance using SHAP, permutation, and impurity decrease methods.

Main Results:

• Models utilizing only three CBC analytes (MCV, RDW, MCH) achieved F1 scores exceeding 90%.
• Model performance plateaued with more than six analytes, indicating efficiency with minimal input.
• MCV and RDW were consistently identified as the most important analytes for WBIT detection.
• Interpretable decision boundaries were visualized in two and three dimensions.

Conclusions:

• Robust WBIT detection is feasible using a small set of CBC analytes.
• Machine learning offers a practical, interpretable, and generalizable solution for WBIT error detection.
• This approach can be implemented across diverse clinical laboratory settings to enhance patient safety.