Abstract 10657: Automated Tracing of Tricuspid Regurgitation Doppler Waveforms

Abstract

Introduction: The definition of pulmonary hypertension relies on the invasive measurement of increased pulmonary arterial pressures (PAPs). The systolic PAP can be non-invasively estimated from the maximal velocity (Vmax) of tricuspid regurgitation (TR) Doppler signals. However, the accuracy of this method is conditioned by the quality of the TR signal. It is also subject to high interobserver variability. To address these issues, we developed an automated, machine-learning framework to analyze TR signals. We hypothesize our method can accurately predict Vmax and velocity time integral (VTI) for any TR signal. Methods: We analyzed 306 TR Doppler signals consisting of high and intermediate signal qualities from 65 patients (median 50 [21, 86] years old). Our method combines a U-Net convolutional neural network with OpenCV, an image processing tool. The U-Net performs semantic image segmentation on an input TR Doppler signal. OpenCV then deduces the Vmax and VTI of the curve. We compared our predictions against expert assessments of the TR signal, Vmax, and VTI. We used 230/56 signals for training/validation to determine the optimal parameters in our U-Net and 20 signals for testing to obtain an unbiased measure of performance. Results: As shown in the figure, our framework accurately predicts Vmax with a bias of 0.06 m/s and 95% limits of agreement of [-0.50, 0.63] m/s, and VTI with a bias of 0.01 m and 95% limits of agreement of [-0.22, 0.23] m. Two signals proved difficult to trace due to noisy images. Our framework takes about 5 seconds to predict Vmax and VTI for each TR signal. Conclusions: The accuracy of the predicted Vmax and VTI is heavily influenced by the quality of the input TR signal. We can solve this issue by training our U-Net on a larger dataset and leveraging other computer-vision techniques or flagging these signals for user intervention. Automation of signal analysis opens the door for more detailed curve analysis on larger clinical cohorts.