Development and evaluation of automated screening algorithms for pre-analytical errors in pharmacokinetic data: a...

Pre-analytical errors such as sample mix-ups can compromise the integrity of pharmacokinetic (PK) data. We simulated vancomycin PK data and introduced two error types: intra-individual time point swap (TS) errors and inter-individual concentration swap (CS) errors. This study developed and validated two automated detection approaches-a run-test-based non-parametric method and a Mahalanobis distance-based method with leave-one-out cross-validation (LOOCV)-for identifying pre-analytical errors in PK datasets. Performance was evaluated using 12,500 simulated profiles comparing the two methods. The distance-based method demonstrated significantly higher error detection rate (79.7% vs. 59.6%, p < 0.001) and greater area under the receiver operating characteristic curve (0.856; 95% confidence interval [CI], 0.816-0.889 vs. 0.574; 95% CI, 0.541-0.610) compared to the run-test, while maintaining reasonable specificity (76.4% vs. 84.0%). For TS errors, the distance-based method achieved higher recall (80.4% vs. 71.9%) and modestly higher F1-score (0.530 vs. 0.515). For CS errors, the distance-based method demonstrated superior precision (86.1% vs. 59.8%) with comparable F1-score (0.572 vs. 0.564). Detection was robust across alternative PopPK models, a structurally different drug (theophylline), training-set contamination (with the minimum covariance determinant estimator recommended for non-curated training), missing data (with sub-vector marginal handling), and sampling-time jitter. Overall, the distance-based method with LOOCV outperformed the conventional run-test across key performance metrics and can serve as a practical quality control tool for ensuring PK data integrity in clinical pharmacology research.