Introduction: The gold standard to diagnose pulmonary hypertension (PH) is invasive right heart catheterization (RHC). However, data from cardiac and 4D flow MRI with modern machine learning models may be able to non-invasively diagnose PH.
Hypothesis: While PH is classically assessed using RHC, we hypothesize that data extracted from cardiac and 4D flow MRI studies used in the context of modern machine learning models can accurately diagnose PH.
Methods: Sixty-four patients with PH (61.3 ± 31.1 years, 41 female) who were diagnosed by RHC and clinical history were prospectively recruited from 2 institutions and gave informed consent. Twenty-four individuals with no clinical features of PH served as controls (64.3 ± 32.4 years , 5 female). All individuals underwent cardiovascular MRI including 4D flow for the measurement of pulmonary 3D flow velocities. MRI data collection included cardiac function, pulmonary artery dimensions, and pulmonary hemodynamics (flow, see Table 1 for parameters, see Figure A for example image). Cases were randomly split into training and testing subsets (i.e. 44 training cases and 44 test cases to evaluate classifier models). Random Forests, Logistic Regression and Gradient Boosted Classifier models from the Python scikit-learn package were trained using 5-fold stratified cross validation. Performance on test cases were evaluated using area-under-curve (AUC) of the receiver operator curves.
Results: Random forests using global function and 4D flow MRI parameters had the best performance (AUC of 0.98 on test cases, Figure B) to automatically classify subjects into PH and control groups. The top 5 most important features identified by the model for classification included MPA diameter/BSA, RVEF, RVEDVi, LVSVi, and LPA net flow index.
Conclusions: Data from 4D flow MRI and modern machine learning methods demonstrate a promising method for diagnosing patients with PH in a non-invasive manner.