Deconstructing Risk: A Machine Learning-Guided Logistic Regression Model for Predicting Early Mucosal Edema Following Prophylactic Ileostomy in Elderly Rectal Cancer Patients

Abstract

Background Early mucosal edema is a common and burdensome complication following prophylactic ileostomy in elderly rectal cancer patients, yet effective tools for its pre-operative prediction are lacking. This study aimed to develop and validate an interpretable, machine learning-guided logistic regression model for this specific outcome. Patients and Methods In this retrospective cohort study, 296 eligible patients were included. The cohort was randomly split into a training set (n=207) for model development and an internal test set (n=89) for validation. A consensus machine learning approach, integrating Least Absolute Shrinkage and Selection Operator (LASSO) regression, stepwise forward logistic regression, and Random Forest with Recursive Feature Elimination (RF-RFE), was employed to select core predictors from comprehensive clinical data. The final predictive model is a standard multivariable logistic regression. Model performance was validated via internal bootstrap resampling, ten-fold cross-validation, and external dataset testing, and assessed by discrimination, calibration, and decision curve analysis, with interpretability enhanced using SHAP (SHapley Additive exPlanations) values. Generalizability was tested in an external cohort (n=40). Results Five variables were consistently selected as core predictors: the Prognostic Nutritional Index (PNI), preoperative neutrophil-to-lymphocyte ratio (NLR), intraoperative fluid intake, abdominal wall opening size, and time to first postoperative bowel movement. The model incorporates early postoperative variables and is intended for early perioperative risk stratification, not solely preoperative prediction. The final multivariable logistic regression model, presented as a nomogram, demonstrated good discriminatory ability in the training (AUC=0.838, 95% CI: 0.706–0.845) and test (AUC=0.826, 95% CI: 0.743–0.910) sets. Internal robustness was supported by ten-fold cross-validation (mean AUC=0.813, 95% CI: 0.680–0.960). The model maintained acceptable performance in the external validation cohort (AUC=0.820), though some miscalibration was noted. Subgroup analysis suggested that a smaller abdominal opening and delayed bowel function were specifically associated with progression to severe edema. Conclusion We developed and validated an interpretable, machine learning-guided logistic regression model for early mucosal edema, incorporating five readily available clinical variables. The model, available as an online nomogram, provides a practical tool for perioperative risk stratification, potentially guiding individualized patient counseling and targeted management strategies to mitigate this complication.