PanEcho: Complete AI-enabled echocardiography interpretation with multi-task deep learning

Abstract

Importance: Echocardiography is a cornerstone of cardiovascular care but relies on expert interpretation and manual reporting from a series of videos. We propose an artificial intelligence (AI) system, PanEcho, to automate echocardiogram interpretation with multi-task deep learning. Objective: To develop and evaluate the accuracy of PanEcho on a comprehensive set of 39 echocardiographic labels and measurements on transthoracic echocardiography (TTE). Design Setting and Participants: This study represents the development and retrospective, multi-site validation of an AI system. PanEcho was developed using a sample of TTE studies conducted at Yale-New Haven Health System (YNHHS) hospitals and clinics from January 2016-June 2022 during routine care. The trained model was internally validated in a temporally distinct YNHHS cohort from July-December 2022, externally validated across four diverse external cohorts, and made publicly available. 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 (MAE) 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. PanEcho performed 18 diagnostic classification tasks with a median AUC of 0.91 (IQR: 0.88-0.93) and estimated 21 echocardiographic parameters with a median normalized MAE of 0.13 (0.10-0.18) in internal validation. For instance, the model accurately estimated left ventricular (LV) ejection fraction (MAE: 4.2% internal; 4.5% external) and detected moderate or higher LV systolic dysfunction (AUC: 0.98 internal; 0.99 external), RV systolic dysfunction (0.93 internal; 0.94 external), and severe aortic stenosis (0.98 internal; 1.00 external). PanEcho maintained excellent performance in limited imaging protocols, performing 15 diagnosis tasks with 0.91 median AUC (IQR: 0.87-0.94) in an abbreviated TTE cohort and 14 tasks with 0.85 median AUC (0.77-0.87) on real-world point-of-care ultrasound acquisitions by non-experts from YNHHS emergency departments. Conclusions and Relevance: We report an AI system that automatically interprets echocardiograms, maintaining high accuracy across geography and time from complete and limited studies. PanEcho may be used as an adjunct reader in echocardiography labs or rapid AI-enabled screening tool in point-of-care settings.

Ähnliche Arbeiten