Date of Award

January 2026

Document Type

Open Access Thesis

Degree Name

Medical Doctor (MD)

Department

Medicine

First Advisor

Michael L. DiLuna

Second Advisor

Aladine A. Elsamadicy

Abstract

Background: Cranial vault remodeling (CVR) for pediatric craniosynostosis is associated with postoperative blood transfusion, yet determinants of transfusion risk remain incompletely defined. Estimating transfusion risk preoperatively could support more personalized blood management planning for this population.Research Aims: To evaluate whether supervised machine learning models can reliably predict postoperative transfusion risk following CVR compared to traditional logistic regression and to identify the most informative predictors of those models. Hypothesis/Question: We hypothesized that ensemble machine learning approaches (e.g. random forest, extreme gradient boosting) would match or outperform logistic regression in predicting transfusion by leveraging nonlinear associations and higher-order interactions across patient, procedural, and perioperative variables. Methods: We performed a retrospective cohort analysis using the 2012-2023 National Surgical Quality Improvement Program Pediatric database, identifying patients aged 2 years or younger who underwent CVR for craniosynostosis. Candidate predictors included demographic variables, comorbidities, preoperative laboratory values, and intraoperative factors. We trained and tested multiple supervised machine learning models (logistic regression, naive Bayes, k-nearest neighbors, decision trees, random forests, and extreme gradient boosting) and assessed performance using measures of discrimination and clinical utility. Results: Among 10,732 patients, 5,781 (53.9%) received a postoperative transfusion. Transfused patients were older (8.53 vs 5.71 months) and heavier (8.26 vs 7.21 kg), more often underwent open procedures (92.5% vs 70.9%), and had longer operative (3.13 vs 1.88 hours) and anesthesia durations (5.07 vs 3.57 hours). The transfusion group had longer hospital length of stay (3.54 vs 2.29 days), higher 30-day adverse events (3.0% vs 1.7%), and higher unplanned readmission (2.9% vs 1.9%) and reoperation (1.9% vs 1.1%). Model performance was similar across machine learning approaches. Extreme gradient boosting achieved the highest discrimination with an area under the receiver operating characteristic curve (AUROC) of 0.794. This was marginally higher than that achieved by random forest (AUROC 0.786) and logistic regression (AUROC 0.783). In the parsimonious multivariable logistic regression model, older age (adjusted odds ratio [aOR] 1.04 per month), longer operative duration (aOR 1.64 per hour), structural central nervous system abnormality (aOR 1.20), and higher preoperative serum creatinine (aOR 5.88) were independently associated with higher transfusion odds. Minimally invasive surgical and hybrid approaches were associated with lower odds of postoperative transfusion compared to open procedures (aOR 0.37 and 0.43, respectively). Random forest variable importance rankings were concordant, prioritizing operative duration and age as the most influential predictors. Scientific Impact: These findings support practical preoperative risk stratification for transfusion in infants undergoing CVR and highlight potentially modifiable targets for perioperative blood management, including consideration of surgical approach and timing of surgery when clinically appropriate. Future work on predictive models may inform counseling and resource planning for patients at high risk of postoperative transfusion.

Comments

This is an Open Access Thesis.

Download
Open Access

This Article is Open Access

Share

COinS