A fully automated deep learning echocardiography workflow achieved an AUC of 0.75 for detecting milder pulmonary hypertension compared to 0.79 for clinical reads (P=0.068).
Case-Control (n=630)
Does a fully automated deep learning workflow in echocardiography reliably assess pulmonary hypertension compared to manual interpretation?
A fully automated deep learning workflow for echocardiography shows comparable reliability to manual interpretation for detecting pulmonary hypertension, potentially improving clinical efficiency.
Absolute Event Rate: 0.75% vs 0.79%
p-value: p=.068
BACKGROUND: Echocardiography is central when assessing pulmonary hypertension (PH), but manual interpretation can be time-consuming and prone to error. RESEARCH QUESTION: Is a fully automated deep learning (DL) workflow in echocardiography reliable when assessing PH? STUDY DESIGN AND METHODS: This study had 2 parts: the first determined the bias and precision of DL reads by using Us2.ai software version 1.4.5 with core laboratory readers as the reference; the second part assessed the ability of DL to discriminate milder PH in patients referred for right heart catheterization (mean pulmonary artery pressure between 20 and 35 mm Hg). The first cohort (case-control) included 213 healthy individuals and 221 patients with pulmonary arterial hypertension. Parameters included peak tricuspid regurgitation velocity (TRV), right ventricular basal diameter, tricuspid annular plane systolic excursion, right atrial area, and right ventricular fractional area change (RVFAC). The referral cohort included 196 patients, with 171 patients having measurable peak TRV signals. Robust measures of bias and precision were reported, and area under the curve (AUC) analysis assessed discrimination. RESULTS: In patients with pulmonary arterial hypertension, mean age was 48 years, 78% were female, and mean pulmonary artery pressure was 52 mm Hg. No significant bias was observed for peak TRV (0.90%; 95% CI, -0.17 to 1.57), right atrial area (1.71%; 95% CI, 0.59 to 3.34), and tricuspid annular plane systolic excursion (1.28%; 95% CI, -0.51 to 3.18), while RVFAC exhibited a significant bias of 11.46% (95% CI, 8.43 to 14.74). For all measurements except RVFAC, robust percentile precision remained below 15%. In the case-control cohort, peak TRV had AUCs of 0.99 and 0.98 for core laboratory and DL reads, respectively. The AUC for PH detection in the referral cohort was 0.79 for clinical laboratory reads and 0.75 for DL reads (P = .068). INTERPRETATION: A fully automated DL workflow for echocardiography in PH is promising and likely to improve efficiency in clinical practice.
Celestin et al. (Tue,) conducted a case-control in Pulmonary Arterial Hypertension (n=630). Fully automated deep learning (DL) workflow in echocardiography vs. Core laboratory readers / clinical laboratory reads was evaluated on Discrimination of milder pulmonary hypertension (AUC) (p=.068). A fully automated deep learning echocardiography workflow achieved an AUC of 0.75 for detecting milder pulmonary hypertension compared to 0.79 for clinical reads (P=0.068).