Abstract
BACKGROUND: Pulmonary hypertension (PH) is a progressive vascular disorder where early diagnosis is critical for improving patient outcomes. While right heart catheterization (RHC) remains the gold standard for diagnosis, its invasive nature often leads to delayed PH detection. This study aimed to develop a machine learning-based predictive model incorporating MRI-derived parameters to facilitate early PH diagnosis. METHODS: In this retrospective study, after data filtering, 323 participants (161 RHC-confirmed PH patients and 162 controls) who underwent cardiac MRI at Zhongnan Hospital, Wuhan University between January 2021 and May 2024 were enrolled for model development, with a 7:3 split for training and internal validation. An additional external validation cohort (48 PH cases and 16 controls) was collected from June 2024 to June 2025. We analyzed 27 MRI parameters reflecting cardiac structure/function, 60 laboratory biomarkers (including NT-proBNP and D-dimer), and basic demographic information (age, sex). Key MRI features were selected via recursive feature elimination (RFE), followed by comparative evaluation of multiple machine learning models (XGBoost, logistic regression, etc.) to identify optimal predictors. SHAP analysis elucidated variable importance, while Random forest selected significant laboratory biomarkers. The final integrated model combined MRI and laboratory predictors, with performance assessed via Receiver Operating Characteristic (ROC) curves, calibration plots, and decision curve analysis (DCA). A nomogram and web-based calculator were developed for clinical implementation. RESULTS: Fifteen MRI parameters showed strong PH association: pulmonary artery diameter (PA), right end-diastolic volumes (REDV), left end-diastolic volumes (LEDV), left end-systolic volumes (LESV), right cardiac index (RCI), right ventricular ejection fraction (RVEF), left atrial anteroposterior diameter (LAAPD), left cardiac output (LCO), right stroke volumes (RSV), left stroke volumes (LSV), left stroke volume index (LSVI), left ventricular lateral wall thickness (LVLWT), left basal interventricular septal thickness (LIVST), and ascending aortic diameters (AAD), descending aortic diameters (DAD). SHAP analysis identified PA, REDV, and LEDV as top predictors. The MRI-derived model demonstrated excellent discriminative ability across all metrics (AUC, precision-recall, specificity-sensitivity). Key laboratory predictors included BUN, γGGT, TBIL, and D-dimer. The combined model achieved AUCs of 0.999 (training), 0.944 (internal validation), and 0.897 (external validation), with excellent calibration. For enhanced clinical utility, we have deployed the developed PH prediction model as a web-based calculator (https://jianghx.shinyapps.io/PH_prediction_MRIIndex/) to facilitate early diagnosis of pulmonary hypertension. CONCLUSION: Our study developed a high-performance PH prediction model integrating cardiac MRI and laboratory biomarkers, demonstrating robust diagnostic accuracy that could enable earlier PH detection while circumventing RHC-related diagnostic delays. CLINICAL TRIAL NUMBER: Not applicable.