Development and validation of a clinical prediction model for postoperative atrial fibrillation after lung cancer surgery: a machine-learning–based study

Figure 1. Flowchart of patient selection, dataset splitting, model development, and validation. A total of 920 patients undergoing lung cancer surgery were screened, of whom 540 were included in the final analysis. Patients were randomly divided into a training cohort (n = 379) and an independent test cohort (n = 161). The training cohort was used for feature selection and model development, and the test cohort was used for model validation and performance evaluation.

Figure 2. Feature selection using least absolute shrinkage and selection operator (LASSO) regression in the training cohort. (A) LASSO coefficient profiles for candidate predictors plotted against −log(λ). As the penalty increased, most coefficients were shrunk toward zero. (B) Ten-fold cross-validation curve for selection of the optimal λ based on binomial deviance. The two vertical dashed lines indicate λ_min (minimum mean cross-validated deviance) and λ_1se (the largest λ within one standard error of the minimum). λ_1se was selected to derive a parsimonious and stable predictor set.

Figure 3. Receiver operating characteristic (ROC) curves of seven machine-learning models in the training and test cohorts. ROC curves illustrate the discriminative performance of logistic regression (LR), k-nearest neighbors (KNN), decision tree (DT), random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM), and neural network (NN) in the training cohort (left) and the independent test cohort (right). The area under the ROC curve (AUC) for each model is shown in the legend.

Figure 4. Calibration curves of the machine-learning models in the training and test cohorts. Calibration curves depict the agreement between predicted probabilities and observed POAF outcomes for each model in the training cohort (A) and test cohort (B). The diagonal line represents perfect calibration (predicted risk equals observed risk). Curves closer to the diagonal indicate better calibration.

Figure 5. Decision curve analysis (DCA) of machine-learning models in the test cohort. Decision curves show the net benefit of each prediction model across a range of threshold probabilities in the independent test cohort. The horizontal line represents the “treat-none” strategy, and the oblique line represents the “treat-all” strategy. A model is considered clinically useful when it provides a higher net benefit than both default strategies over clinically relevant thresholds.

Figure 6. Nomogram for individualized prediction of postoperative atrial fibrillation. A nomogram integrating age, hypertension, surgical approach, postoperative pain score, marital status, and education level to estimate an individual's risk of POAF. For each predictor, points are assigned according to its value; the total points correspond to the patient's predicted probability of POAF.