Abstract 4370353: Prognostic Value of RV Endocardial Radiomics in Patients with Pulmonary Hypertension

Background: Right ventricular (RV) function is highly predictive of mortality in patients with pulmonary hypertension (PH). RV ejection fraction and RV end-systolic stroke volume index significantly associate with mortality but less is known about higher-order cardiac magnetic resonance (CMR) features including Radiomic features. Hypothesis: We hypothesize that radiomic features in an XGBoost model can improve predictions of one-year mortality in patients with PH. Methods: Patients with four-chamber Cine cardiac magnetic resonance imaging were identified in the OSU CMR/RHC PH registry. Standard radiomic features were extracted from RV endocardium regions of interest (ROI) using Matlab (R2024b). Contrast was adjusted in grayscale images using histogram equalization (histeq). Shape, intensity and texture-based features were extracted across the cardiac cycle (102 features/frame). To capture time-series changes, radiomic features were assessed at end-diastole (ED), end-systole (ES), and averaged during systole (sys) and diastole (dia), yielding 408 features per subject. Features with no variance and highly correlated features (|r|>0.9) were filtered out (n = 305). Three predictive models were evaluated: 1) Clinical CMR features, 2) Radiomics features alone, and 3) Combined CMR and Radiomics features. XGBoost, a gradient boosting algorithm, was used to predict mortality. Model performance was assessed using the area under the receiver operating characteristic curve (AUC). The relative importance of each feature was also assessed across all models using Shapley Additive exPlanations (SHAP) values. Results: Radiomic features were extracted in 97 participants (age: 55±15 years) with an average RV ejection fraction of 41±14% and 25 deaths at 1 year. The radiomics model showed better accuracy (0.89, 95%CI: 0.67-0.99) and AUC (0.6) compared to clinical CMR features alone (Accuracy: 0.79, 95%CI: 0.54-0.94, AUC: 0.93). Combined clinical CMR and Radiomics features did not improve the accuracy. Important features in the clinical CMR model included RV end-diastolic volume index, LV end-systolic volume index, and RV ejection fraction ( Figure 1A ). Top 3 features in the Radiomics model included a mix of shape-based, gray-level cooccurrence matrix, and first order features ( Figure 1B ). Conclusions: RV endocardial radiomics features improved prediction of death in a small mixed cohort of patients with PH over clinical CMR measures alone.