Identifying Aortic Stenosis With a Single Parasternal Long-Axis Video Using Deep Learning

Abstract

The accurate diagnosis of aortic stenosis (AS) involves both the acquisition of cardiac ultrasound images and the interpretation of these images by skilled personnel.1Otto C.M. Nishimura R.A. Bonow R.O. et al.2020 ACC/AHA guideline for the management of patients with valvular heart disease: a report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines.Circulation. 2021; 143: e72-e227PubMed Google Scholar Access to such specialty care, however, may not be possible in many parts of the world, and regular echocardiographic studies can be expensive. Nevertheless, AS is a progressive disorder, and follow-up echocardiographic studies are recommended for patients with valvular disease.2Carabello B.A. Evaluation and management of patients with aortic stenosis.Circulation. 2002; 105: 1746-1750Crossref PubMed Scopus (101) Google Scholar A quick and accurate detection method, which minimizes the need for specialized clinical interpretation, would make AS screening more accessible in settings where access to clinical specialists is limited. Deep learning (DL) forms a platform for the automated evaluation of echocardiographic data. Indeed, DL models have been constructed for a variety of tasks ranging from view classification to disease diagnosis.3Madani A. Arnaout R. Mofrad M. et al.Fast and accurate view classification of echocardiograms using deep learning.NPJ Digit Med. 2018; 1: 6Crossref PubMed Scopus (254) Google Scholar,4Shad R. Quach N. Fong R. et al.Predicting post-operative right ventricular failure using video-based deep learning.Nat Comm. 2021; 12: 5192Crossref PubMed Scopus (17) Google Scholar In the case of AS, Huang et al.5Huang Z, Long G, Wessler B, et al. A new semi-supervised learning bench-mark for classifying view and diagnosing aortic stenosis from echocardio-grams. Proceedings of Machine Learning Research; volume 149, 2021. Proceedings of the 6th Machine Learning for Healthcare Conference.Google Scholar reported a DL model that classifies patients as having no AS versus mild or severe AS, using all views obtained during a routine echocardiographic study.5Huang Z, Long G, Wessler B, et al. A new semi-supervised learning bench-mark for classifying view and diagnosing aortic stenosis from echocardio-grams. Proceedings of Machine Learning Research; volume 149, 2021. Proceedings of the 6th Machine Learning for Healthcare Conference.Google Scholar The method was developed and evaluated on a small data set (260 patients) and achieved a 90% accuracy in the classification task. We hypothesized that a single parasternal long-axis (PLAX) view could be used to identify severe AS. To construct a DL model for identifying severe AS using a single PLAX view, we identified all echocardiographic studies performed from 2001 to 2019 at Massachusetts General Hospital and selected studies where the mean transvalvular gradient or aortic valve area (AVA) was reported by a level III trained echocardiographer. This resulted in a total number of 28,734 studies from 16,066 patients, where all studies had a reported a mean transaortic valve pressure. Out of these studies, 15,041 studies, arising from 8,749 patients, had measured AVAs. Data preprocessing and summary statistics of patients used to develop each model are shown in the Supplemental Material. We trained and tested 3 DL models: (1) MP: a model to identify when the mean transvalvular pressure was >40 mm Hg using all 28,734 studies; (2) MA: a model to identify when the AVA was <1 cm2 using 15,041 studies; and (3) MS: a model to identify whether either the transvalvular pressure is above 40 mm Hg or the AVA is below 1 cm2. Each model examines a single PLAX view (movie) as input and outputs a prediction. The model training procedure is described in the Supplemental Material. Table 1 details the discriminatory ability of each model. Model MP has better discriminatory ability relative to the other models, likely because more data were available for training this model. Sensitivity-specificity curves as well as positive and negative predictive values at different prevalence levels are shown in Figure 1A. To gauge how this model would perform in the general population, estimates of the prevalence of severe AS are needed. Several studies across different patient cohorts have estimated the global prevalence of severe AS to be between 3% and 4% in patients over 75 years of age.6Eveborn G.W. Schirmer H. Heggelund G. et al.The evolving epidemiology of valvular aortic stenosis. the Tromsø study.Heart. 2013; 99: 396-400Crossref PubMed Scopus (383) Google Scholar At this prevalence level, the negative predictive value of all 3 models is more than 98%, with an 80% sensitivity.Table 1Discriminatory ability of different modelsModelTest set sizeArea under the curve (SD)Mp5,7910.88 ± 0.01Ma3,0750.78 ± 0.01MS3,0750.79 ± 0.01 Open table in a new tab To understand what data within a PLAX view most influence model predictions, we calculated saliency maps. Saliency map analysis is an illustrative way to reveal what regions of an image most influence model decision-making.7Simonyan K. Vedaldi A. Zisserman A. Deep inside convolutional networks: visualising image classification models and saliency maps.CoRR. 2014; (abs/1312.6034)Google Scholar The method calculates a scalar “saliency value” for each pixel in an image, and the resulting matrix is called a saliency map. The larger the saliency value, the more important that pixel is for the model arriving at its prediction. Examples of saliency analyses are shown in Figure 1B, which displays randomly selected studies from 3 patients. In each case, pixels around the aortic valve are clearly highlighted compared to the background in both models. This demonstrates that DL models align appropriately with the relevant anatomic feature. Screening for AS in patients with relevant risk factors and/or clinical exam findings suggestive of aortic valvular disease remains a mainstay of clinical care. In this study, we developed several DL models to identify severe AS patients using only a single PLAX video. All models have good discriminatory ability and have high negative predictive value at a prevalence level expected for patients over 75 years old, suggesting that these methods could be used to effectively rule out severe AS in this cohort. The DL algorithms can be applied in an automated manner, in the context of a point-of-care ultrasound study, to facilitate screening of patients with severe AS. In addition, this screening algorithm can help sonographers and echocardiographers prioritize what studies to focus on, that is, studies that the model identifies as severe AS warrant a more thorough evaluation of the aortic valve. We have made the model generally available at https://github.com/mit-ccrg/AS_PLAX. Download .pdf (.57 MB) Help with pdf files Supplemental Material