Beyond the abdomen: an interpretable machine learning model for predicting postoperative ileus in non-abdominal surgery

Figures

• 

  Figure 1. Feature selection and correlation analysis. (A) LASSO coefficient profiles of candidate variables. (B) Selection of the optimal lambda (λλ) in the LASSO model using 10-fold cross-validation. (C) Dynamic feature importance evolution during the Boruta algorithm process. (D) Final feature importance ranking by Boruta; green box plots indicate confirmed predictors. (E) Spearman correlation heat map of the seven selected variables.

• 

  Figure 2. Feature importance ranking across eight machine learning algorithms. Bar charts display the relative contribution of predictors for: (A) Random Forest; (B) Naïve Bayes; (C) Logistic Regression; (D) LGBM; (E) Decision Tree; (F) SVM; (G) KNN; (H) XGBoost. Color intensity corresponds to the importance score.

• 

  Figure 3. Performance evaluation in training (top row), internal testing (middle row), and external validation (bottom row) cohorts. (A–C) ROC curves comparing the discriminatory ability of the eight models. (D–F) Calibration plots assessing agreement between predicted probabilities and observed frequencies. (G–I) Decision curve analysis (DCA) illustrating the net clinical benefit of the models.

• 

  Figure 4. Model performance metrics and statistical comparison in the training cohort. (A) Comparison of AUC values among the eight models. (B) Heat map of comprehensive performance metrics. (C) Heatmap of P-values from pairwise DeLong tests comparing the ROC curves of the 8 machine learning models. The values in the cells represent the raw P-values. Statistical significance was evaluated following Bonferroni correction for multiple comparisons. The RF model’s superiority remained robust and highly significant against all other models.

• 

  Figure 5. Five-fold cross-validation of the random forest model. ROC curves showing the stability of the model in the: (A) Training Cohort; (B) Internal Testing Cohort; (C) External Validation Cohort. Solid black lines represent the mean ROC, while dashed colored lines represent individual folds.

• 

  Figure 6. SHAP analysis for random forest model interpretation. (A) mean absolute SHAP values ranking global feature importance. (B) SHAP summary plot showing the impact of feature values on prediction (color indicates feature value; x-axis indicates impact). (C) Waterfall plot illustrating feature contributions for a high-risk patient. (D) Waterfall plot illustrating feature contributions for a low-risk patient.