A Machine Learning Approach to Protect Patients from Blood Tube Mix-Ups
summary
A deployed system detects wrong‑blood‑in‑tube (WBIT) errors by training binary classifiers on multi‑analyte delta features computed from paired CBCs, with synthetic WIBITs created via matchit temporal/department matching to overwrite results and label outcomes. Models are built with tidymodels, tuned for high positive predictive value using a downsampled holdout at 1‑in‑3k–5k prevalence to estimate real‑world false positives, and planned for production via vetiver and Posit Connect with an API to integrate lab middleware and the EMR.
AI-written summary · about
chapters (10)
- 0:00Preventing patient harm with ML
- 2:51Current measures for blood tubes
- 4:19Multi-analyte ML approach
- 7:56Simulating WIBIT errors
- 9:02Training ML models
- 10:27Clinical workflow constraints
- 13:15Adjusting for low prevalence
- 15:04Optimizing alert thresholds
- 17:14Trade-offs and plans
- 18:39Conclusion and impact
related talks
22mReimaging Healthcare
20mDiagnostic Medical application of ML
13mMEDebiaser: A Human-AI Feedback System for Mitigating Bias in Multi-label Medical Image Classific...
35mResponsible implementation of Machine Learning models for life-and-death decision making
39mBest Practices for Prototyping ML Models for Healthcare
39mBridging health inequalities through machine learning
05mDBTB INT Eric Williams r
38mRealtime prediction of C-section risk for laboring mothers