Automated Assessment of Right Atrial Pressure From Ultrasound Videos Using Machine Learning.

Background: Early recognition of volume overload is essential for heart failure patients. Volume overload can often be easily treated if caught early but causes significant morbidity if unrecognized and allowed to progress. Intravascular volume status can be assessed by ultrasound-based estimation of right atrial pressure (RAP), but the availability of this diagnostic modality is limited by the need for experienced physicians to accurately interpret these scans.
Objectives: We sought to evaluate whether machine learning can accurately estimate echocardiogram-measured RAP.
Methods: We developed fully automated deep learning models for identifying inferior vena cava scans with rapid inspiration in echocardiogram studies and estimating RAP from those scans. The RAP estimation model was trained and evaluated using 15,828 ultrasound videos of the inferior vena cava and coupled cardiologist-assessed RAP estimates as well as 319 RAP measurements from right heart catheterization.
Results: Our model agreed with cardiologist estimates 80.3% of the time (area under the receiver-operating characteristic of 0.844) in a test data set, at the upper end of interoperator agreement rates found in the literature of 70 to 75%. Our model's RAP estimates were statistically indistinguishable from cardiologists' ultrasound-based RAP estimates ( P  = 0.98) when compared against the gold standard of right heart catheterization RAP measurements in a subset of patients. Our model also generalized well to an external data set of echocardiograms from a different institution (area under the receiver-operating characteristic of 0.854 compared to cardiologist RAP estimates).
Conclusions: Machine learning is capable of accurately and robustly interpreting RAP from echocardiogram videos. This algorithm could be used to perform automated assessments of intravascular volume status.
Competing Interests: Dr Tison has received research grants from 10.13039/100004313General Electric. Dr Avram has received speaker fees from Abbott, Boston Scientific, Boehringer-Ingelheim, and Novartis. All other authors have reported that they have no relationships relevant to the contents of this paper to disclose. Dr Abu-Mostafa received research grants from three internal Caltech research endowments: the 10.13039/100023001Merkin Institute for Translational Research (#13520291), The Sensing to Intelligence Initiative (#13520296), and the Gates-Grubstake fund (#101170). Dr Tison received support from the 10.13039/100000002National Institutes of Health (Grants NHLBI K23HL135274, R56HL161475, and DP2HL174046). Dr Padmanabhan received support from the 10.13039/100000002National Institutes of Health (Grant NHLBI K08HL157700), Michael Antonov Charitable Foundation, and 10.13039/100009540Frank A. Campini Foundation. Dr Avram received support from the Fonds de la recherche en santé du Québec (Grant 312758), by 10.13039/100007631CIFAR, by the Montreal Heart Institute Research Centre, the 10.13039/501100012651Montreal Heart Institute Foundation, and by the Des Groseillers-Bérard Research Chair. The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.
(© 2024 The Authors.)