Predicting Post-surgery Discharge Time in Pediatric Patients Using Machine Learning.

BACKGROUND

Prolonged hospital stays after pediatric surgeries, such as tonsillectomy and adenoidectomy, pose significant concerns regarding cost and patient care. Dissecting the determinants of extended hospitalization is crucial for optimizing postoperative care and resource allocation.

OBJECTIVE

This study aims to utilize machine learning (ML) techniques to predict post-surgery discharge times in pediatric patients and identify key variables influencing hospital stays.

METHODS

The study analyzed data from 423 children who underwent tonsillectomy and/or adenoidectomy at the IRCCS Istituto Giannina Gaslini, Genoa, Italy. Variables included demographic factors, anesthesia-related details, and postoperative events. Preprocessing involved handling missing values, detecting outliers, and converting categorical variables to numerical classes. Univariate statistical analyses identified features correlated with discharge time. Four ML algorithms-Random Forest (RF), Logistic Regression, RUSBoost, and AdaBoost-were trained and evaluated using stratified 10-fold cross-validation.

RESULTS

Significant predictors of delayed discharge included postoperative nausea and vomiting (PONV), continuous infusion of dexmedetomidine, fentanyl use, pain during discharge, and extubation time. The best-performing model, AdaBoost, demonstrated high accuracy and reliable prediction capabilities, with strong performance metrics across all evaluation criteria.

CONCLUSION

ML models can effectively predict discharge times and highlight critical factors impacting prolonged hospitalization. These insights can enhance postoperative care strategies and resource management in pediatric surgical settings. Future research should explore integrating these predictive models into clinical practice for real-time decision support.