Development and external validation of an explainable machine learning model using routine perioperative clinical variables for postoperative pulmonary complication risk stratification after hepatectomy for primary liver cancer

Figures

• 

  Figure 1. Study flowchart. AUC, area under the curve; DCA, decision-curve analysis; LASSO, least absolute shrinkage and selection operator; LightGBM, Light Gradient Boosting Machine; MICE, multiple imputation by chained equations; NPV, negative predictive value; PLC, primary liver cancer; PPC, postoperative pulmonary complication; PR-AUC, area under the precision-recall curve; SHAP, Shapley Additive Explanations; XGBoost, Extreme Gradient Boosting.

• 

  Figure 2. Feature selection process using LASSO regression and RF-RFE algorithms. (A) LASSO coefficient profiles of the candidate clinical features. (B) Selection of the optimal tuning parameter (λ) in the LASSO model using 10-fold cross-validation. (C) Selection of the optimal number of features using the RF-RFE algorithm. (D) Venn diagram demonstrating the overlap between variables selected by the LASSO and RF-RFE algorithms. ASA, American Society of Anesthesiologists; BMI, body mass index; LASSO, least absolute shrinkage and selection operator; RF-RFE, random forest recursive feature elimination.

• 

  Figure 3. ROC and PR curves of the machine learning models. (A) ROC curves in the training set. (B) ROC curves in the internal testing set. (C) PR curves in the training set. (D) PR curves in the internal testing set. AUC, area under the curve; DT, decision tree; KNN, k-nearest neighbors; LightGBM, Light Gradient Boosting Machine; LR, logistic regression; NB, Naïve Bayes; PNI, prognostic nutritional index; PR, precision-recall; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine; XGBoost, Extreme Gradient Boosting.

• 

  Figure 4. Comprehensive evaluation metrics of the eight machine learning models. (A) Heatmap of model performance metrics in the training set. (B) Heatmap of model performance metrics in the internal testing set. AUC, area under the curve; DT, decision tree; KNN, k-nearest neighbors; LightGBM, Light Gradient Boosting Machine; LR, logistic regression; NB, Naïve Bayes; NPV, negative predictive value; PR-AUC, area under the precision-recall curve; RF, random forest; SVM, support vector machine; XGBoost, Extreme Gradient Boosting.

• 

  Figure 5. Model calibration and clinical utility assessment. Calibration curves assessing the agreement between predicted probabilities and actual observed outcomes for PPCs in the (A) training set and (B) internal testing set. DCA demonstrating the net clinical benefit of the predictive models across a wide range of probability thresholds in the (C) training set and (D) internal testing set. DCA, decision-curve analysis; DT, decision tree; KNN, k-nearest neighbors; LightGBM, Light Gradient Boosting Machine; LR, logistic regression; NB, Naïve Bayes; PPC, postoperative pulmonary complication; RF, random forest; SVM, support vector machine; XGBoost, Extreme Gradient Boosting.

• 

  Figure 6. Confusion matrix of the eight models for PPC classification in the training and internal testing cohorts. DT, decision tree; KNN, k-nearest neighbors; LightGBM, Light Gradient Boosting Machine; LR, logistic regression; NB, Naïve Bayes; PPC, postoperative pulmonary complication; RF, random forest; SVM, support vector machine; XGBoost, Extreme Gradient Boosting.

• 

  Figure 7. Performance and clinical utility of the NB model in the external validation cohort. (A) ROC curve and (B) PR curve. (C) Calibration curve. (D) DCA. AUC, area under the curve; DCA, decision-curve analysis; NB, Naïve Bayes; PR, precision-recall; PR-AUC, area under the precision-recall curve; ROC, receiver operating characteristic.

• 

  Figure 8. Model interpretability and individualized risk prediction utilizing SHAP analysis for the NB model. (A) Feature importance ranking based on mean absolute SHAP values. (B) SHAP summary plot illustrating the distribution and directional impact of feature values on the model output. (C,D) SHAP waterfall plots providing individualized explanations of model predictions. These plots quantify the contribution of each clinical feature for (C) a representative high-risk patient and (D) a representative low-risk patient. Yellow bars denote features that increase the likelihood of PPCs, whereas red bars denote protective features that decrease the risk. ASA, American Society of Anesthesiologists; NB, Naïve Bayes; PNI, prognostic nutritional index; PPC, postoperative pulmonary complication; SHAP, Shapley Additive Explanations.

• 

  Figure 9. Interface of the web-based risk calculator. Based on the final NB model, the application enables rapid bedside postoperative risk stratification by generating an individualized PPC risk probability, corresponding risk classification, and SHAP-based feature contribution plot. ASA, American Society of Anesthesiologists; NB, Naïve Bayes; PPC, postoperative pulmonary complication; SHAP, Shapley Additive Explanations.