Application of Machine Learning Models Integrating Clinical and Echocardiography in the Prediction of Mean Pulmonary Artery Pressure Grading

Background As a progressive cardiopulmonary disorder, pulmonary hypertension (PH) necessitates precise assessment of mean pulmonary arterial pressure (mPAP) for clinical staging, treatment planning, and prognostic evaluation. Methods We retrospectively included patients who underwent right heart catheterization (RHC) at our institution between January 2017 and October 2025. The cohort was divided temporally into a training cohort (January 2017 to December 2023) and a validation cohort (January 2024 to October 2025). Echocardiographic parameters and clinical data were collected. A four-category label was constructed based on mPAP grading (0–20 mmHg, 21–35 mmHg, 36–45 mmHg, > 45 mmHg). Key features were selected using Lasso combined with the Boruta method. The Synthetic Minority Over-sampling Technique (SMOTE) balanced the training cohort sample distribution. Ultimately, eight machine learning (ML) models were constructed and their performance evaluated. Model performance was assessed using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1 score, and Brier score. Feature importance for predictive models was interpreted using SHapley Additive exPlanations (SHAP) values. Results A total of 495 patients were included in model construction. Six features were selected from 29 variables for model training: 6-Minute Walk Distance (6MWD), Eccentricity Index (EI), Left Ventricular Diameter (LVD), Right Ventricular Diameter (RVD), Tricuspid Annular Plane Systolic Excursion/Pulmonary Artery Systolic Pressure (TAPSE/PASP), and PASP. Among all ML models, the Naive Bayes model achieved the highest classification accuracy, with an AUC of 0.886, accuracy of 0.736, Brier score of 0.106, and F1 score of 0.736. Its AUC within the training cohort reached 0.894. Furthermore, the mean AUC values for different mPAP classifications were 0.994, 0.878, 0.779, and 0.892, respectively. SHAP value analysis confirmed that TAPSE/PASP was the primary predictive feature for mPAP classification, followed by PASP and EI. These three features demonstrated consistent performance across all subgroups. Conclusions The non-invasive predictive model developed in this study provides a reliable tool for the precise classification of mPAP in PH patients, thereby assisting clinicians in reducing reliance on invasive RHC.