A Multianalyte Machine Learning Model to Detect Wrong Blood in Complete Blood Count Tube Errors in a Pediatric Setting

Background Multianalyte machine learning (ML) models can potentially identify previously undetectable wrong blood in tube (WBIT) errors, improving upon current single-analyte delta check methodology. However, WBIT detection model performance has not been assessed in a real-world, low-prevalence context. To estimate real-world positive predictive values, we propose a methodology to assess WBIT detection models by evaluating the impact of missing data and by using a “low prevalence” validation data set. Methods We trained a range of model specifications using various predictors in a pediatric setting. We assessed the top-performing model on a modified, “low prevalence” validation data set across a range of probability thresholds. Model performance was also compared to a pre-positive patient identification (pre-PPID) dataset. Results An Extreme Gradient Boosting (XGBoost) model with minimal preprocessing performed the best for both complete blood count with differential white cell count (CBC with Diff) tests (accuracy 0.9715) and complete blood count without differential white cell count (CBC without Diff) tests (accuracy 0.9647). Assessment on a downsampled, “low prevalence” validation data set resulted in estimated positive predictive values ranging from 0.01 to 0.67 (CBC with Diff) and 0.01 to 0.75 (CBC without Diff), depending on the probability threshold chosen. A comparison of prospective performance to PPID data demonstrated a large decrease in estimated WBIT errors. Conclusions We find that ML models can accurately predict WBITs in a primarily pediatric setting. Evaluating model performance across a range of probability thresholds minimizes the number of false positives while still providing added safety benefits. The decrease in estimated WBITS post-PPID implementation shows the potential safety benefits of a WBIT model for hospitals not using PPID when collecting laboratory specimens.