Severe aortic stenosis detection by deep learning applied to echocardiography.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Early diagnosis of aortic stenosis (AS) is crucial for preventing severe patient morbidity and mortality.
• Current diagnostic methods require skilled examination and Doppler imaging, limiting point-of-care accessibility.
• A novel deep learning approach aims to simplify AS detection.

Purpose of the Study:

• To develop and validate a deep learning model for identifying severe AS.
• To assess the model's performance using 2D echocardiography videos without Doppler imaging.
• To determine the model's utility for point-of-care ultrasonography screening.

Main Methods:

• An ensemble of 3D convolutional neural networks was trained on 5257 transthoracic echocardiography studies (2016-2020).
• Self-supervised contrastive pretraining was used for label-efficient model development.
• The model was validated on temporally distinct (2040 studies) and geographically distinct (4226 and 3072 studies) cohorts.

Main Results:

• The deep learning model achieved an AUROC of 0.978 in the primary test set.
• High diagnostic performance was maintained in external validation cohorts (AUROC 0.952 and 0.942).
• Saliency maps confirmed model interpretability, highlighting key cardiac structures.

Conclusions:

• An automated approach for severe AS detection was developed and externally validated.
• The model utilizes single-view 2D echocardiography, enhancing its potential for point-of-care screening.
• This AI-driven tool could significantly improve early AS detection and management.