Development and validation of a machine learning model for accurate detection of wrong blood in tube errors in...

To develop and validate a machine-learning model based on routinely available biochemical and hematological parameters for detecting wrong blood in tube (WBIT) errors in hospitalized patients.

A retrospective multicenter study including one internal cohort (IC) and two external validation cohorts (EVC, EVC2). The IC was balanced (50 % correct, 50 % WBIT; 25 % real, 25 % simulated), while EVC (n=800) and EVC2 (n=460) represented more realistic scenarios (95 % correct, 5 % WBIT; equally distributed between real and simulated). Parameters present in ≥ 95 % of requests were selected, and their normalized variation from the immediately preceding result was calculated. The IC was divided into a training set (IC-TS, n=324) and an internal validation set (IC-VS, n=108). Feature selection was refined with Elastic Net before training an XGBoost model. Performance was assessed in IC-VS, EVC, and EVC2. For benchmarking, the model's discriminative ability was also compared with a multivariate Mahalanobis-based approach and with univariate delta checks within IC-TS/IC-VS.

Sixteen of 25 candidate variables were retained. The model achieved ROC-AUC values of 0.98-0.99 and PR-AUC values of 0.93-0.99 across all validation cohorts. Recalibration improved positive predictive value and net benefit by reducing false positives, with a slight decrease in sensitivity, although all values remained ≥90 %. Specificities ranged from 98 to 99 %. The model consistently outperformed both the multivariate Mahalanobis approach and univariate delta checks within the internal cohort.

This machine-learning model, leveraging widely available routine laboratory parameters, shows strong potential for integration into clinical workflows, enhancing WBIT detection and improving patient safety.