WOFAPS 2025 8th World Congress of Pediatric Surgery

İncinur Genişol ¹, Asya Eylem Boztaş ¹, İlker Özgür Koska ², Arzu Şencan ³

¹ University of Health Sciences, Dr. Behçet Uz Pediatric Diseases and Surgery Training and Research Hospital, Department of Pediatric Surgery, Izmir
² University of Health Sciences, Izmir Faculty of Medicine, Dr. Behçet Uz Pediatric Diseases and Surgery Training and Researc Hospital, Department of Radiology, Izmir
³ University of Health Sciences, Izmir Faculty of Medicine, Dr. Behçet Uz Pediatric Diseases and Surgery Training and Research Hospital, Department of Pediatric Surgery, Izmir

Aim:This study aimed to develop machine learning models to predict surgical timing and discharge duration in patients with infantile hypertrophic pyloric stenosis(IHPS), using clinical, biochemical, and ultrasonographic data available at admission.

Materials and Methods:A retrospective analysis was conducted on 55 IHPS cases operated between 2015-2025. Demographic, biochemical(blood gas, electrolytes, bilirubin, and urinalysis) and ultrasonographic data(pyloric wall thickness, transverse diameter, length, pyloric index, and pyloric volume) were evaluated. Three datasets were created:(1)demographic and imaging variables, (2)demographic and laboratory variables, and (3)all variables combined.Target outcomes were defined as undergoing surgery within 2 days of admission and being discharged within 3 days postoperatively. Class imbalance was addressed using SMOTE and class weighting. For SVM and RF, four model types were created:standard, SMOTE-enhanced, class weight-adjusted, an ensemble including extra trees. Models were trained using 4-fold cross-validation(10 repeats). Evaluation metrics included F1 score, accuracy, sensitivity, and ROC, AUC. LASSO was used for most predictebl selection.

Results:In predicting surgical timing, the best performance was achieved with the SMOTE-enhanced SVM model using the combined dataset.For discharge prediction, the best results were obtained with the SMOTE-enhanced RF model using the second dataset and the SVM model using the combined dataset. LASSO analysis identified pyloric wall thickness and blood gas pH as the most influential features for early surgery(F1:0.64, Accuracy0.59, Sensitivity %62, Specifity%67), while pyloric transverse diameter and urine specific gravity were most predictive for early discharge(F1:0.64, Accuracy0.59, Sensitivity %62, Specifity %67)(Table-1).

Conclusion:The best performance was seen with SVM models using all data types.LASSO analysis identified pyloric wall thickness and blood gas pH as key predictors for early surgery, and pyloric transverse diameter and urinary specific gravity for discharge. These results support the utility of machine learning in guiding decision-making in IHPS managemenent.