Comparison of logistic regression and random forest models for predicting pegaspargase-associated acute pancreatitis in pediatric patients with acute lymphoblastic leukemia

Objective To develop and compare logistic regression (LR) and random forest (RF) models for predicting pegaspargase (PEG-ASP)-associated acute pancreatitis (AAP) in children with acute lymphoblastic leukemia (ALL). Methods This retrospective two-center study enrolled children with ALL who received PEG-ASP between January 2020 and April 2025. AAP was diagnosed according to the Chinese national guideline for acute pancreatitis, and children who received at least one PEG-ASP dose were eligible. Candidate predictors were screened using univariate analysis, entered into a multivariable LR model, and an RF model was built in parallel. Both models were developed within the training data and assessed by repeated stratified 5-fold cross-validation and a held-out test set. Discrimination, calibration, and clinical usefulness were compared. Results Of 714 children with ALL who received PEG-ASP, 34 (4.8%) developed AAP. After excluding records with incomplete predictor data, 142 patients (31 with AAP and 111 without) comprised the model-development dataset. In the multivariable LR model, a lower total protein (TP) level (OR 0.875, 95% CI 0.803–0.955; P = 0.003) and a higher absolute neutrophil count (ANC) (OR 2.947, 95% CI 1.338–6.488; P = 0.007) were independently associated with AAP. Both models ranked TP and ANC as the two most influential predictors. In repeated cross-validation, discrimination was modest and slightly better for LR (AUC 0.765, 95% CI 0.649–0.861) than for RF (AUC 0.708, 95% CI 0.582–0.827). The two models performed similarly on the held-out test set, and the LR model’s calibration was acceptable. Conclusion Lower TP and higher ANC before PEG-ASP administration were associated with AAP in children with ALL. The LR and RF models showed only modest and broadly comparable performance in this small, two-center cohort and should be regarded as preliminary tools that require external validation before clinical use.