Introduction: Reliable prediction of 1-year mortality in newly diagnosed pulmonary arterial hypertension (PAH) is important because the first year after diagnosis is a key window for follow-up treatment planning and transplant assessment. Established clinical risk approaches may not fully capture the complementary information provided by demographic, laboratory, and imaging data. This study therefore evaluated whether multimodal machine-learning models could improve 1-year mortality prediction beyond established clinical risk approaches.
Methods: This retrospective study used ASPIRE registry data collected between 2008 and 2025. 1-year mortality was modelled using demographic and laboratory data, together with right heart catheterisation (RHC) and cardiovascular magnetic resonance (CMR) data. REVEAL, ESC/ERS, COMPERA and WHO functional class were benchmarked using 4-fold cross-validation. Multiple machine-learning models were developed using RHC-only, CMR-only, and combined RHC+CMR feature groups. Class imbalance was addressed using the Synthetic Minority Over-sampling Technique for Nominal and Continuous features (SMOTE-NC) and class weighting. Backward recursive feature elimination was applied to derive reduced models.
Results: Among established clinical risk approaches, REVEAL showed the best performance (AUC=0.699), outperforming COMPERA (AUC=0.571), WHO functional class (AUC=0.540) and ESC/ERS (AUC=0.522). The support vector machine with SMOTE-NC showed the highest performance among the evaluated models. In pairwise comparisons, the CMR-only and combined RHC+CMR models outperformed their paired REVEAL benchmarks (AUC 0.739 vs 0.701, Δ=0.037, p<0.001; and 0.735 vs 0.692, Δ=0.043, p<0.001, respectively), whereas the RHC-only model showed slightly lower performance than REVEAL (0.667 vs 0.683, Δ=-0.016, p=0.004). Model reduction further improved predictive performance, with the reduced 9-feature model derived from the combined RHC+CMR feature group achieving a cross-validation AUC of 0.797 (SD=0.075).
Conclusion: Multimodal machine-learning models, particularly those incorporating CMR, improve 1-year mortality prediction over established clinical risk approaches in newly diagnosed PAH. They can also be reduced to a smaller, potentially implementable feature set with improved predictive performance.