Development and Validation of a Risk Model to Predict Intraoperative Blood Transfusion.

Summary

Using 816,618 surgical cases across two health systems, the TRANSFUSE model (24 preoperative variables) achieved an AUC of 0.93 for predicting intraoperative pRBC transfusion and outperformed a widely used score. Internal and external validation confirmed generalizability, and predictive values improved in higher-risk operations.

Key Findings

• Model trained and validated on 816,618 surgeries with AUC 0.93 (95% CI 0.92–0.93).
• Included 24 preoperative predictors (e.g., ASA status, INR, redo/emergency surgery, duration ≥120 min, surgical complexity, anemia, liver disease, thrombocytopenia, surgery type).
• Outperformed the Transfusion Risk Understanding Scoring Tool (AUC 0.64) and matched or exceeded 3 ML-derived scores; NPV 99.7% overall.

Clinical Implications

Incorporate TRANSFUSE into preoperative workflows to right-size crossmatch orders (especially in high-risk procedures), embed in EHR decision support, and align with PBM protocols to minimize non-transfused units.

Limitations

• Overall PPV is modest (8.9%) due to low base rate; calibration may vary across institutions and case mix.
• Observational registry data may include unmeasured confounding and coding variability.

Future Directions

Prospective impact analyses to quantify blood product savings, adverse event reduction, and cost-effectiveness; calibration/transportability studies and user-centered EHR integration.