New AI Model Outperforms Standard Echocardiograms in Heart Disease Detection

A recently developed multiview deep neural network (DNN), a type of AI algorithm, has outperformed standard single-view DNNs in detecting signs of heart disease through 3D imaging, according to a recent study.1

Standard echocardiogram devices typically operate via 2D visual images of 3D cardiac anatomy. They take hundreds of individual slices, or views, of a beating heart to detect any abnormalities in cardiovascular structure or function. Clinicians must then reassemble these slices to construct a 3D mental model, allowing direct analysis and assessment of the characteristics determining the presence or absence of disease.1,2

Most of these devices, however, even when paired with AI, take only a single view of the heart. In the present study, Geoffrey Tison, MD, MPH, co-director of the University of California San Francisco (UCSF) Center for Biosignal Research, and Joshua Barrios, PhD, an assistant professor in the UCSF Division of Cardiology, programmed a DNN capable of taking images from multiple angles at once, thereby allowing it to search for multiple potential indicators of cardiovascular disease at once.1,2

“Our AI architecture was motivated by the way that cardiologists read echoes – we tend to read them by looking at multiple views at the same time to come up with a single diagnosis,” Tison told HCPLive in an exclusive interview. “Most AI models with echo have looked at only a single view, and derived a diagnosis from that view, so our architecture aims to allow the model to learn across multiple views and thus arrive at a diagnosis that is appropriate across all of those views.”

Tison, Barrios, and colleagues developed the DNN model to accept inputs from multiple views, integrating information through dedicated AI layers. The architecture utilized a mid-fusion approach, combining features from each view to allow the network to integrate inter-view information.1

This architecture was applied to 3 main demonstration echo tasks, including the identification of left or right ventricular (LV/RV) abnormalities, the identification of substantial valvular regurgitation, and the identification of diastolic dysfunction. The first 2 were considered “standard” echo tasks, given that accurate manual interpretation typically involves corroboration from >1 view, while the third was a “novel” echo task that physicians cannot typically interpret using non-doppler, brightness mode (B-mode) echo.1

The team collected a series of previous echo studies from patients at UCSF to compare the DNN against single-view AI devices trained on each demonstration task individually. For the LV/RV abnormality task, a cohort of 41,790 echo studies from 20,504 patients was collected, with a prevalence of LV/RV abnormality of 24.5%. The multiview DNN achieved an area under the receiver operating characteristic curve (AUC) of 0.907 (95% CI, 0.9-0.914) with a sensitivity and specificity of 0.81 and 0.84, respectively, in this task. A total of 3 single-view DNNs were incorporated for this task – the best of these exhibited an AUC of 0.851 (95% CI, 0.841-0.861).1

Similarly, in the diastolic dysfunction task, the multiview DNN scanned 11,411 echo studies from 6643 patients for an AUC of 0.836 (95% CI, 0.821-0.851) with a sensitivity and specificity of -0.76. The best single-view DNN performance had an AUC of 0.749 (95% CI, 0.73-0.767). The novel valve regurgitation task followed this trend, with a cohort of 27,692 studies from 18,573 patients included. The multiview DNN achieved an AUC of 0.904 (95% CI, 0.892-0.915) with a specificity and sensitivity of -83%. The best single-view DNN performance was an AUC of 0.836 (95% CI, 0.821-0.852).1

Based on these data, Tison and Barrios believe that this device is far more effective at detecting cardiovascular disease than standard echocardiograms, with or without AI implementation. However, they also note that similar studies must be conducted across other institutions, given that the multiview DNN examines scans based on their labels, which intrinsically differ between centers.1

“There’s a real challenge here related to differences in practice and differences in labels,” Barrios told HCPLive in an exclusive interview. “The way that labels are generated is different across institutions, and this is a general problem for AI. We’re wrestling with realities of historical data since we’re not doing this prospectively.”

Editors’ Note: Tison reports disclosures with Viz.ai, Prolaio, and MyoKardia. Barrios reports no relevant disclosures.

References

1. Barrios JP, Ansari MU, Olgin JE, et al. Multiview deep learning improves detection of major cardiac conditions from echocardiography. Nature Cardiovascular Research. 2026;5(3):234-245. doi:10.1038/s44161-026-00786-7

2. Krigel M. This Multiview AI Imaging Can Improve Cardiac Diagnostic Accuracy. UCSF Research. March 16, 2026. Accessed May 21, 2026. https://www.ucsf.edu/news/2026/03/431611/multiview-ai-imaging-can-improve-cardiac-diagnostic-accuracy