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

Area of Science:

• Clinical Laboratory Science
• Machine Learning in Healthcare
• Patient Safety

Background:

• Current methods for detecting blood draw errors are limited.
• Multianalyte machine learning (ML) models offer potential for improved detection of wrong blood in tube (WBIT) errors.
• Real-world performance of WBIT detection models in low-prevalence settings is not well-established.

Purpose of the Study:

• To assess the real-world performance of ML models for WBIT error detection.
• To propose a methodology for evaluating WBIT detection models in low-prevalence contexts.
• To compare ML model performance against existing pre-positive patient identification (pre-PPID) methods.

Main Methods:

• Trained various ML model specifications using diverse predictors in a pediatric cohort.
• Assessed the top-performing Extreme Gradient Boosting (XGBoost) model on a "low prevalence" validation dataset.
• Evaluated model performance across different probability thresholds and compared it to pre-PPID data.

Main Results:

• The XGBoost model achieved high accuracy for both CBC with Diff (0.9715) and CBC without Diff (0.9647) tests.
• Estimated positive predictive values for WBIT detection ranged from 0.01 to 0.75, depending on the probability threshold.
• Prospective performance comparison showed a significant reduction in estimated WBIT errors compared to PPID data.

Conclusions:

• ML models can accurately predict WBITs in pediatric laboratory settings.
• Optimizing probability thresholds balances false positive reduction with safety benefits.
• WBIT ML models offer potential safety advantages for laboratories not utilizing PPID.