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

Background: Wrong blood in tube (WBIT) is a critical pre-analytical error in laboratory medicine in which a blood sample is mislabeled with the wrong patient identity. These errors are often undetected due to the limitations of current detection strategies (e.g., delta checks). Methods: We evaluated Random Forest models for WBIT detection and conducted a detailed analyte importance analysis. In total, 799,721 samples from a German tertiary care center were analyzed and filtered for applicability. Model input features were derived by pairing consecutive same-patient samples for non-WBIT cases, simulating WBIT by pairing samples from different patients, and computing per-analyte first-order differences for each pair. We exhaustively searched all subsets of nine CBC analytes and evaluated models using F1 score, AUC, sensitivity, and PPV. Analyte importance was assessed via SHAP, permutation, and impurity decrease. Results: Models using as few as three analytes (MCV, RDW, MCH) reached F1 scores above 90%, with performance plateauing beyond six analytes. MCV and RDW were consistently top-ranked. Two-dimensional and three-dimensional visualizations revealed interpretable decision boundaries. Conclusions: Findings demonstrate that robust WBIT detection is achievable using a minimal subset of CBC analytes, offering a practical, interpretable, and broadly generalizable ML-based solution suitable for diverse clinical environments.