Development and nested cross-validation of machine learning models for predicting transfusion requirements in pediatric craniosynostosis surgery
Journal of Plastic, Reconstructive & Aesthetic Surgery · 2026-07 · 118 · 209-220
DOI: 10.1016/j.bjps.2026.05.033 ↗Kurz gefasst
Validierungsstudie in der Zeitschrift Journal of Plastic, Reconstructive & Aesthetic Surgery, 2026-07. Verfasst von Jiawen Deng und 6 weiteren Autor:innen. Beteiligte Einrichtung: Temerty Faculty of Medicine, University of Toronto, Toronto, ON, Canada. Hauptthemen laut MEDLINE: Craniosynostoses, Blood Transfusion, Machine Learning, Craniotomy. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 43 Literaturangaben.
Abstract (Originalsprache)
Blood transfusion is common during pediatric craniosynostosis surgery; however, transfusion volumes and use of cell salvage systems can vary considerably. To support preoperative transfusion planning in patients undergoing craniectomy for craniosynostosis, we developed and internally validated machine learning (ML) models to predict perioperative transfusion risks and transfusion volumes in this patient population. Data from eligible patients ≤36 months of age from the NSQIP® Pediatric dataset were included for modeling. A classifier pipeline was developed to predict probabilities associated with transfusion risk, and a probabilistic/quantile regression pipeline was developed to predict the median transfusion volume (in mL/kg) with interquartile range. Performance of the modeling pipelines was evaluated using nested cross-validation with 5-fold outer loops. In total, 8045 patients were included with 4091 (50.9%) receiving perioperative transfusion. For the classification pipeline, the optimal model configuration was a LightGBM classifier. The modeling pipeline achieved a mean area under the receiver operating characteristic curve (AUC-ROC) of 0.77 (95% confidence interval [CI] 0.75-0.78) and mean Brier score of 0.20 (95% CI 0.19-0.20) across nested cross-validation. For the regression stage, the optimal model configuration was a probabilistic NGBoost regressor. The modeling pipeline achieved an average median absolute error of 6.86 mL/kg (95% CI 6.61-7.11), with a mean IQR coverage of 53% (95% CI 52-53%). These models may support individualized blood product preparation and selective use of cell salvage systems. External validation and further real-world testing are needed prior to the clinical implementation of this tool.
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Temerty Faculty of Medicine, University of Toronto, Toronto, ON, Canada.
- Mayo Clinic Alix School of Medicine, Mayo Clinic, Rochester, MN, USA.
- Department of Internal Medicine, Mayo Clinic, Rochester, MN, USA.
- Department of Anesthesiology, Mayo Clinic, Rochester, MN, USA.
- Department of Neurosurgery, Mayo Clinic, Rochester, MN, USA.
- Division of Plastic Surgery, Department of Surgery, Mayo Clinic, Rochester, MN, USA.
Zitieren
APA
Deng, J., Gharavi, A., Heybati, K., Haile, D., Ahn, E. S., Mardini, S., & Gibreel, W. (2026). Development and nested cross-validation of machine learning models for predicting transfusion requirements in pediatric craniosynostosis surgery. Journal of Plastic, Reconstructive & Aesthetic Surgery, 118, 209-220. https://doi.org/10.1016/j.bjps.2026.05.033
Vancouver
Deng J, Gharavi A, Heybati K, Haile D, Ahn ES, Mardini S, et al. Development and nested cross-validation of machine learning models for predicting transfusion requirements in pediatric craniosynostosis surgery. Journal of Plastic, Reconstructive & Aesthetic Surgery. 2026;118:209-220. doi:10.1016/j.bjps.2026.05.033
BibTeX
@article{Deng2026,
author = {Deng, Jiawen and Gharavi, Aidin and Heybati, Kiyan and Haile, Dawit and Ahn, Edward S and Mardini, Samir and Gibreel, Waleed},
title = {Development and nested cross-validation of machine learning models for predicting transfusion requirements in pediatric craniosynostosis surgery},
journal = {Journal of Plastic, Reconstructive & Aesthetic Surgery},
year = {2026},
volume = {118},
pages = {209--220},
doi = {10.1016/j.bjps.2026.05.033}
}
EUVIMED