Ultrasound-Independent Rapid Triage in Pediatric Appendicitis via Explainable AI

Abstract

Ultrasonography (USG) is central to the diagnosis of pediatric appendicitis, but operator dependency and limited access can delay triage. This study aimed to develop an explainable machine learning model that estimates the USG-assisted clinical diagnosis of appendicitis from pre-imaging clinical and laboratory data alone. Ten algorithms were trained on 780 patients from the public Regensburg Pediatric Appendicitis Dataset after removing ultrasonographic, post-diagnostic, and label-defining variables. In five repetitions of stratified 80/20 splitting, models were tuned and internally validated by nested cross-validation with fold-local preprocessing, and each holdout set was used once for evaluation. Discrimination was assessed with the area under the receiver operating characteristic curve (AUC) and probabilistic accuracy with the Brier score; the primary model was explained with SHapley Additive exPlanations (SHAP). The Generalized Additive Model (GAM) achieved an AUC of 80.5% ± 2.2% (mean ± standard deviation) and a Brier score of 0.182 ± 0.012 in the repeated holdout (nested cross-validation AUC 77.3% ± 1.0%), comparable to linear models. Its negative predictive value was 66.2%, and it underestimated risk at the lowest predicted probabilities. White blood cell count (WBC), peritonitis, and C-reactive protein (CRP) accounted for 31.3% of the total mean absolute SHAP contribution, followed by age and body mass index. A USG-independent GAM provided moderate discrimination and interpretable risk estimates that may support early risk stratification before imaging. Its limited negative predictive value precludes its use for ruling out appendicitis, and external and prospective validation is required before clinical use.