Machine learning constructs a diagnostic prediction model for gangrenous perforation of acute appendicitis in elderly patients

BACKGROUND: As life expectancy rises and elderly populations grow, acute appendicitis incidence increases, often manifesting with nonspecific symptoms that challenge diagnosis. This study applied machine learning techniques to build a predictive model for gangrenous perforation, examining clinical features and risk factors in elderly patients with acute appendicitis. METHODS: We conducted a retrospective analysis of elderly patients undergoing laparoscopic appendectomy for acute appendicitis at The Second Affiliated Hospital of Kunming Medical University, China, from June 2021 to January 2024 (n = 251). Patients were classified into gangrenous perforation (n = 69) and non-gangrenous (n = 182) groups. For descriptive reporting, patients were randomly split into training (70%) and test (30%) sets (fixed seed); univariate comparisons and feature selection were restricted to the training set. Candidate predictors were selected by LASSO and further reduced by logistic regression. Model development and performance estimation used repeated nested cross-validation (outer 5-fold ×10 repeats; inner 5-fold; metric: AUC; seed = 123). LR_3var, XGBoost, SVM, and RF were evaluated using out-of-fold AUC, calibration, and decision curve analysis. RESULTS: LASSO retained nine candidates; logistic regression identified three final predictors (WBC, CRP, albumin). Mean outer-fold AUCs were 0.756 ± 0.080 (LR_3var), 0.757 ± 0.078 (XGBoost), 0.757 ± 0.081 (SVM_linear), and 0.728 ± 0.083 (RF). Multivariable models outperformed single predictors; calibration and decision curves indicated acceptable calibration and potential net benefit. CONCLUSIONS: Using routine laboratory markers, multivariable models showed good discrimination for gangrenous perforation under repeated nested cross-validation. LR_3var was selected for nomogram-based risk stratification due to interpretability and practicality.