Exploratory early-warning modeling of in-hospital mechanical intestinal obstruction after kasai portoenterostomy in children with biliary atresia

Abstract

Mechanical intestinal obstruction during hospitalization after Kasai portoenterostomy for biliary atresia (BA) is uncommon but clinically important. Because event numbers are limited and early manifestations overlap with postoperative ileus, this retrospective study aimed to identify perioperative early-warning signals and construct an exploratory, interpretable risk-stratification model while minimizing information leakage and overfitting. This single-center retrospective chart-review study included 322 children with BA who underwent Kasai portoenterostomy between 2014 and 2024. The primary outcome was in-hospital mechanical intestinal obstruction, defined by compatible clinical features plus radiographic evidence and distinguished from postoperative ileus. Six algorithms were evaluated using repeated stratified fivefold cross-validation (10 repeats; 50 outer folds) with fold-specific imputation, Boruta-lite feature selection, training-fold-only SMOTE oversampling in the primary analysis, a no-SMOTE sensitivity analysis, and standardization when required. The incidence of in-hospital mechanical intestinal obstruction was 2.5% (8/322). Logistic regression showed the strongest internal performance, with a fold-level mean AUROC of 0.9867 ± 0.0156 and a fold-level mean AUPRC of 0.7637 ± 0.2430. Longer abdominal distension, shorter daily passive mobilization time, and intraoperative fluid volume were the most consistent early-warning signals. PR-curve analysis, F1-score reporting, calibration assessment, decision-curve analysis, and the no-SMOTE sensitivity analysis were used to evaluate model behavior under class imbalance. Because only eight events were available, these estimates remain exploratory and likely optimistic. In this single-center retrospective cohort, prolonged abdominal distension and reduced daily passive mobilization mainly represented early postoperative red flags, whereas intraoperative fluid volume represented a perioperative signal. The Logistic regression model should be regarded as an exploratory red-flag aggregator for low-threshold monitoring after Kasai portoenterostomy, not as an externally validated clinical decision rule.