Original Investigation
AI in Medicine
Complete AI-Enabled Echocardiography Interpretation With Multitask Deep Learning
Gregory Holste, Evangelos K. Oikonomou, Márton Tokodi, et al
JAMA Published Online: June 23, 2025
doi: 10.1001/jama.2025.8731
Question Can artificial intelligence (AI) fully automate echocardiogram interpretation?
Findings This study reports the development and validation of an automated AI system for echocardiogram analysis, PanEcho, that performed 18 diagnostic classification tasks with a median area under the receiver operating characteristic curve of 0.91 and 21 echocardiographic parameter estimation tasks with a median normalized mean absolute error of 0.13.
Meaning An AI system can automate complete echocardiogram interpretation with high accuracy, potentially accelerating workflows and enabling rapid cardiovascular health screening in point-of-care settings with limited access to trained experts.
Abstract
Importance Echocardiography is a cornerstone of cardiovascular care, but relies on expert interpretation and manual reporting from a series of videos. An artificial intelligence (AI) system, PanEcho, has been proposed to automate echocardiogram interpretation with multitask deep learning.
Objective To develop and evaluate the accuracy of an AI system on a comprehensive set of 39 labels and measurements on transthoracic echocardiography (TTE).
Design, Setting, and Participants This study represents the development and retrospective, multisite validation of an AI system. PanEcho was developed using TTE studies conducted at Yale New Haven Health System (YNHHS) hospitals and clinics from January 2016 to June 2022 during routine care. The model was internally validated in a temporally distinct YNHHS cohort from July to December 2022, externally validated across 4 diverse external cohorts, and publicly released.
Main Outcomes and Measures The primary outcome was the area under the receiver operating characteristic curve (AUC) for diagnostic classification tasks and mean absolute error for parameter estimation tasks, comparing AI predictions with the assessment of the interpreting cardiologist.
Results This study included 1.2 million echocardiographic videos from 32 265 TTE studies of 24 405 patients across YNHHS hospitals and clinics. The AI system performed 18 diagnostic classification tasks with a median (IQR) AUC of 0.91 (0.88-0.93) and estimated 21 echocardiographic parameters with a median (IQR) normalized mean absolute error of 0.13 (0.10-0.18) in internal validation. For instance, the model accurately estimated left ventricular ejection fraction (mean absolute error: 4.2% internal; 4.5% external) and detected moderate or worse left ventricular systolic dysfunction (AUC: 0.98 internal; 0.99 external), right ventricular systolic dysfunction (AUC: 0.93 internal; 0.94 external), and severe aortic stenosis (AUC: 0.98 internal; 1.00 external). The AI system maintained excellent performance in limited imaging protocols, performing 15 diagnosis tasks with a median (IQR) AUC of 0.91 (0.87-0.94) in an abbreviated TTE cohort and 14 tasks with a median (IQR) AUC of 0.85 (0.77-0.87) on real-world point-of-care ultrasonography acquisitions from YNHHS emergency departments.
Conclusions and Relevance In this study, an AI system that automatically interprets echocardiograms maintained high accuracy across geography and time from complete and limited studies. This AI system may be used as an adjunct reader in echocardiography laboratories or AI-enabled screening tool in point-of-care settings following prospective evaluation in the respective clinical workflows.
