Non-invasive estimation of hemodynamic parameters in pulmonary hypertension with a deep learning approach integrating all B-mode cine loops in an echocardiographic exam

Abstract Background Pulmonary hypertension (PH) is heterogeneous with treatment strategy dependent on the underlying cause and disease severity. Hemodynamic parameters measured through right heart catheterization (RHC) is the gold standard for the diagnosis. However, RHC is invasive and is associated with a certain level of risk. Therefore, a non-invasive alternative would be clinically valuable. Echocardiography is acquired routinely as the first-line non-invasive investigation. In the current practice, tricuspid regurgitation velocity (TRVmax) derived from an echocardiographic exam offers a non-invasive estimation of PH likelihood. However, the approach is limited to only subjects with tricuspid regurgitation. In contrast, B-mode images from echocardiography provide comprehensive information about the heart, and thus serve as a valuable source for deriving hemodynamic parameter estimations. Purpose We seek to investigate whether hemodynamic parameters can be estimated non-invasively using a deep learning approach, integrating multi-view B-mode echocardiographic cine loops. Methods The study is based on a retrospective analysis of 833 subjects with suspected PH identified from the ASPIRE research database, where both echocardiography and RHC data were available for analysis. A convolutional neural network was built to predict each of the key hemodynamic parameters, including mean pulmonary artery pressure (mPAP), pulmonary vascular resistance (PVR), and pulmonary artery wedge pressure (PAWP). The model input consists of all the B-mode cine loops from an echocardiographic exam, which are an arbitrary number of videos from multiple views, unannotated with view names. Attention weights were used to identify cine loops that are considered relevant by the model. Results For mPAP, PVR, and PAWP, the model predictions correlated to the RHC-ground truth with Pearson Correlation Coefficients (PCC) of 0.70, 0.68, and 0.54, respectively. Thirty-eight exams in the test set had no TRVmax, but our method managed to estimate hemodynamic parameters for these exams. Pre-capillary PH in the absence of left heart disease, defined as mPAP 20 mmHg, PVR 2 Wood units, and PAWP ≤ 15 mmHg, could be predicted with an accuracy of 0.68, sensitivity of 0.89, and specificity of 0.42. According to the attention weights, the mPAP model identified the A4C view to be especially relevant for mPAP prediction. Conclusion Our results demonstrate the feasibility of estimating hemodynamic parameters non-invasively through deep learning models, integrating all B-mode cine loops of a cardiac ultrasound exam, achieving a moderate correlation to RHC measurements.