Research on the construction of an artificial intelligence model for the early diagnosis of pediatric acute appendicitis: a single-center retrospective study based on complete blood count.

BACKGROUND

Acute appendicitis is the most common pediatric surgical emergency, but is difficult to diagnose. This study aimed to develop and validate a machine-learning model for the early diagnosis of acute appendicitis in children, thereby improving diagnostic accuracy, reducing missed and misdiagnosis, and assisting clinicians in timely therapeutic decision-making.

METHODS

A retrospective cohort of 2,379 children admitted to Children’s Hospital of Chongqing Medical University between January 2010 and January 2025 with acute abdominal pain as the chief complaint was analyzed. After applying predefined inclusion and exclusion criteria, 961 patients were assigned to the appendicitis group (diagnosis confirmed by ultrasound combined with clinical signs or by post-operative histopathology) and 1,418 to the non-appendicitis group (including undifferentiated abdominal pain, gastroenteritis, functional pain, etc.). The cohort was randomly split 7:3 into a training set (n = 1,665) and a test set (n = 714). Univariate analysis was used to identify risk factors, which were then entered into a multivariable LR model. Internal validity was assessed by 1,000-bootstrap resampling and external validity by the held-out test set. ROC curve, calibration plots, and DCA were used to evaluate discriminative performance and clinical utility. To further improve performance, eight machine-learning algorithms (RF, LightGBM, GBDT, SVM, XGBoost, LR, NaïveBayes, and KNN) were trained. Model interpretability was provided by SHAP analysis.

RESULTS

Twelve indicators were incorporated into the model construction, including Pain duration, sex, fever, nausea, PNR, NEUT, CRP, LYMPH, PLR, AISI, MLR, and LMR. The conventional multivariable LR model showed modest performance with limited generalizability. Among the eight ML models, RF achieved the highest discriminative ability and best generalization (test-set AUC 0.828). SHAP summary plots identified the top five contributors in the RF model as Pain duration, PNR, NEUT, CRP, and fever.

CONCLUSION

A risk-prediction model integrating clinical signs and routine complete blood count indices can accurately identify pediatric acute appendicitis. The RF-based model offers superior and generalizable diagnostic performance, providing clinicians with an objective tool for early decision-making.