Machine Learning Prediction of Hospital Stay in Pediatric Typhoid Intestinal Perforation: Pilot Study

Typhoid intestinal perforation (TIP) remains a significant cause of pediatric morbidity in resource-limited settings, with prolonged hospitalization straining healthcare resources. Accurate prediction of hospital stay duration could optimize resource allocation and clinical decision-making, yet no TIP-specific prediction models exist. The objective of this research was to develop and validate a machine learning models for predicting prolonged hospital stay (>8 d) in pediatric TIP patients using readily available admission parameters, comparing logistic regression and random forest approaches for clinical decision support.

A retrospective analysis of 90 pediatric TIP patients from University of Ilorin Teaching Hospital, Nigeria (2009-2020) was conducted. Five clinically accessible features were evaluated: age, admission weight, serum potassium, serum sodium, and packed cell volume. Prolonged stay was defined as length of stay (LoS) >8 d (median value). We developed and validated two machine learning classifiers: logistic regression and random forest. Dataset augmentation using synthetic minority oversampling techniques expanded the cohort to 1500 instances while preserving original data distributions. Both classifiers were developed using repeated stratified k-fold cross-validation (10 repeats × 5 folds). A user-friendly graphical interface was created for point-of-care risk stratification.

The original dataset included 90 patients with mean LoS of 8.41 ± 2.20 d. The augmented dataset maintained realistic distributions with mean LoS of 8.41 ± 2.19 d. The logistic regression model demonstrated good discrimination with cross-validation mean area under the curve (AUC) of 0.715 ± 0.066 (95% confidence interval: 0.649-0.781). The random forest model achieved excellent discrimination with mean AUC of 0.887 ± 0.042 (95% confidence interval: 0.845-0.929). Both models showed stable performance across validation folds. Serum potassium emerged as the dominant predictor (34.1% feature importance), followed by packed cell volume (22.7%), sodium (18.6%), weight (14.3%), and age (10.3%). A Python-based graphical interface enables real-time risk stratification at the point of care.

This study demonstrates the feasibility of developing clinically deployable machine learning tools for pediatric TIP risk stratification. The random forest model achieved excellent discrimination (AUC 0.887), while the logistic regression model provides good performance (AUC 0.715) with maximum interpretability. Data augmentation successfully addressed sample size limitations inherent in single-center surgical studies. Serum potassium's dominant prognostic role warrants protocolized monitoring and correction strategies. External validation in independent cohorts is essential before clinical deployment. The study provides a functional framework and user interface for TIP outcome prediction in resource-limited settings.