Abstract / Summary
Abstract Background Among chronic kidney disease (CKD) patients referred for right heart catheterization (RHC) due to clinical suspicion of pulmonary hypertension (PH), PH is a highly prevalent complication and contributes to increased cardiac mortality. This study aimed to develop and validate a machine learning (ML)-based predictive model for PH risk stratification and to guide RHC referral triage in this high-risk population. Methods In this study, a total of 393 consecutive CKD patients who underwent RHC at Nanjing First Hospital between January 2014 and November 2023 were enrolled as the development cohort. All patients were randomly divided into a training set (70%) and an internal validation set (30%) using stratified random sampling (stratified by PH status and CKD stage). An independent prospective temporal external validation cohort of 400 consecutive CKD patients referred for RHC at the same center between December 2023 and March 2026 was enrolled with identical inclusion/exclusion criteria, and no model retraining or parameter tuning was performed on the validation data. After performing least absolute shrinkage and selection operator regression analysis to select predictive factors, we developed five ML models: logistic regression, k-nearest neighbors, support vector machine, gradient boosting machine, and extreme gradient boosting (XGB). To determine the optimal model, we evaluated various performance metrics including area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score. Finally, the Shapley additive explanation was applied to elucidate each predictor’s impact on the optimal model, and decision curve analysis was performed to assess the clinical applicability. Robustness of the model and sampling strategy was validated via 100 repeated random splits. Results Six variables including left atrial diameter (LAD), blood urea nitrogen (BUN), right atrial transverse diameter (RATD), red blood cell distribution width (RDW), right ventricle diameter (RVD), and pericardial effusion (PE) were identified as independent predictive factors for prevalent PH in CKD patients undergoing RHC. Among the five ML models, the XGB model exhibited the best predictive performance, with an AUC of 0.852 (95% CI: 0.805–0.899) in the training set and 0.824 (95% CI: 0.669–0.874) in the internal validation set. In the temporal external validation cohort, the XGB model maintained robust performance with an AUC of 0.802 (95% CI: 0.751–0.853), significantly outperforming conventional echocardiography (AUC 0.802 vs. 0.713, Delong test Z = 2.87, P = 0.004). The model showed excellent calibration (Hosmer-Lemeshow χ²=7.34, df = 8, P = 0.501) and consistent predictive performance across CKD stages, age groups and genders. Decision curve analysis confirmed superior clinical utility compared to “treat all” and “treat none” strategies across most clinically relevant thresholds. Robustness validation demonstrated stable model performance independent of cohort partitioning, with a median AUC of 0.821 (IQR: 0.803–0.837) across 100 splits. Conclusion The XGB model incorporating LAD, BUN, RATD, RDW, RVD and PE exhibits excellent and robust predictive performance in both internal and temporal external validation cohorts, and significantly outperforms conventional echocardiography. It may serve as a reliable non-invasive risk stratification tool for prioritizing RHC evaluation in high-risk CKD patients with clinical suspicion of PH.