You are on page 1of 1

2409

JACC March 7, 2023


Volume 81, Issue 8, suppl A

Spotlight on Special Topics


SMART-AS: A NOVEL ARTIFICIAL INTELLIGENCE TOOL TO DETECT SEVERE AORTIC STENOSIS
FROM ELECTROCARDIOGRAPHIC IMAGES
Poster Contributions
Poster Hall_Hall F
Monday, March 6, 2023, 12:45 p.m.-1:30 p.m.

Session Title: Spotlight on Special Topics: Spotlight on Special Topics: Innovation, Digital Health, and Technology 18
Abstract Category: 60. Spotlight on Special Topics: Innovation, Digital Health, and Technology
Presentation Number: 1812-033

Authors: Veer Sangha, Evangelos K. Oikonomou, Akshay Khunte, Kartikeye Gupta, Bobak Mortazavi, Rohan Khera, Yale University, New
Haven, CT, USA
Background: Early detection of severe aortic stenosis (AS) is critical for timely intervention to prevent morbidity and mortality. We sought
to develop a layout-independent artificial intelligence (AI) model that can identify severe AS directly from electrocardiographic (ECG)
images.
Methods: We identified 475,968 ECGs with paired echocardiographic readings within 15 days of the index ECG (age 68±10 years,
n=213,859 women), including n=6,988 (1.5%) with severe AS to train a convolutional neural network (CNN) model predicting severe AS.
This was subsequently combined with the patientʼs age, sex, and predicted ejection fraction (EF) from our previously validated AI-ECG
EF model to build an ensemble model. In addition to internal validation, we validated our model on 200 ECG images from Lake Regional
Hospital in rural Missouri (n=8, 4% with severe AS).
Results: In the internal testing set, the model had an AUROC of 0.84 (95%CI, 0.81-0.87). For the ensemble model AUROC was 0.87
(95%CI, 0.84-0.89). The model demonstrated consistent performance across different lead layouts. In external validation, the model
performed comparably (AUROC 0.85 and 0.87, respectively). GradCAM analysis showed localization of discriminative signal to the anterior
precordial leads.
Conclusion: A novel AI-ECG model can reliably detect severe AS directly from ECG images without the need for raw signal data. This
approach may enable point-of-care screening of severe AS across diverse settings with minimal resource utilization.

You might also like