Abstract 14250: Semi-automated Analysis of Tricuspid Regurgitation Doppler Profile for Detection and Evaluation of Pulmonary Hypertension

Abstract

Introduction: Non-invasive detection of pulmonary hypertension relies on the tricuspid regurgitant (TR) Doppler maximal velocity (Vmax). When the TR signal is incomplete, Vmax must be estimated by extrapolation, which is prone to error. Automatic extrapolation could improve robustness but requires an in-depth understanding of complete TR shape. Semi-automated analysis was used to test the hypothesis that all waveforms can be expressed with a small set of curve parameters. Methods: We retrospectively analyzed complete continuous Doppler TR signals from 76 patients (median 50 [39, 56] years-old, 77% female) with pulmonary arterial hypertension in whom echocardiography and invasive hemodynamics were available within 3 months. Semi-automated analysis divided Doppler images into heartbeats (three signals per patient) using the ECG and recognized segmented waveforms for shape analysis. Multiple linear regression modeling assessed the independent association between waveform features (Vmax, velocity-time integral, curve width, skewness (t1,t2), kurtosis and time scale) and age, sex, pulmonary vascular resistance (PVR), TR severity, right ventricular (RV) end-diastolic area index (size), RV strain (function) and right atrial volume index. Results: TR waveforms were non-Gaussian with varying skewness, width and kurtosis. Figure summarizes the covariates associated with TR waveform features. Vmax was directly related to PVR and RV strain, but not to TR severity. The RR-normalized width was correlated to heart rate, TR severity, RV strain and size (all p<0.0001). Kurtosis was not only associated with heart rate but also with PVR, TR severity, RV strain and size (all p<0.01). Conclusions: The non-Gaussian TR Doppler curve shape is influenced by TR severity, RV function and size. The high variability in shape limits the ability to extrapolate signals in the presence of an incomplete envelope. Machine learning algorithms may be valuable to address this issue.