A machine learning ultrasomics model predicted left ventricular (LV) remodeling, which was significantly associated with major adverse cardiovascular events (MACE) in both cardiac point-of-care ultrasound (POCUS) and using images from a high-end ultrasound system, a new study shows.
Quincy A. Hathaway, PhD, of West Virginia University, Naveena Yanamala, PhD, of Rutgers Robert Wood Johnson Medical School, and colleagues reported these findings in a manuscript published Monday online and in the Dec. 8 issue of the Journal of the American College of Cardiology.
POCUS is one of the more noteworthy innovations in echocardiography in recent years, showing the ability to improve medical decision-making. However, a major drawback of this system has been the difficulty of performing gated morphometric measurements on the device’s small screen.
The authors posited that artificial intelligence might help with automated image acquisition and measurements, but most of these platforms “have relied on deep-learning techniques that are inherently a ‘black box,’ causing uncertainty regarding how they operate and, ultimately, how they come to decisions,” they wrote.
Radiomic approaches – mathematical and statistical methods to extract features from radiologic images – have recently been developed. These approaches evaluate images and provide interpretable spatial information on pixel or voxel distribution and patterns. This method allows not only visual interpretation but also quantifies patterns not visible to the human eye.
Extending this concept to ultrasound images is known as “ultrasomics,” which also provides useful measurements, but “little is known about the role of modern ultrasomics approaches for extracting cardiac textural properties, and no studies to date have explored ultrasomics in cardiac POCUS images,” the authors wrote.
Therefore, the authors studied whether ultrasomics can provide rapid, automated assessment of LV structure and function without requiring manual measurements.
They developed machine-learning models using cardiac ultrasound images from 1,915 subjects in three clinical cohorts: an expert-annotated cardiac POCUS registry (n=943), a prospective POCUS cohort for external validation (n=275), and a prospective external validation on high-end ultrasound systems (n=484). Also, the authors assessed ultrasomics’ biological basis in a mouse model, 10 with wild-type diabetes and eight with type 2 diabetes, at 3 and 25 weeks, and correlated the ultrasomics features with the histopathological features of hypertrophy.
The primary outcome was LV remodeling, for which the authors used a broad definition, including any change in the heart’s size, shape and function. This was used to conduct large-scale association mining between the cardiac ultrasomics features and echocardiographic markers of cardiac remodeling, which included structural changes (LV hypertrophy, LV dilation or left atrial dilation) or functional changes (left ventricular ejection fraction <50%, wall-motion abnormalities, or both). A key secondary outcome was the association of ultrasomics probability score with MACE in the external validation cohorts.
The ultrasomics model predicted LV remodeling in both POCUS (area under the curve [AUC]: 0.78; 95% confidence interval [CI]: 0.68-0.88) and high-end ultrasound (AUC: 0.79; 95% CI: 0.73-0.86). The ultrasomics model was also significantly associated with MACE in both cohorts (POCUS: hazard ratio [HR] 0.76, 95% CI 3.27-19.9, p≤0.0001; high-end ultrasound: HR 2.35, 95% CI 1.23-4.47, p=0.0008).
On a multivariate analysis, the ultrasomics probability score was an independent echocardiographic predictor of MACE in the high-end ultrasound cohort (HR 8.53, 95% CI 4.75-32.1, p=0.0003).
In the mouse model, cardiomyocyte hypertrophy was positively correlated with two ultrasomics biomarkers (R2 = 0.57 and 0.52, Q < 0.05).
These results suggest that cardiac ultrasound-based biomarkers might help develop machine-learning models “that provide an expert-level assessment of LV structure and function,” the authors concluded.
Thomas H. Marwick, MBBS, PHD, MPH, of the University of Melbourne, Australia, gave high praise for the study in an accompanying editorial.
“This paper sets a high standard in this field, using well defined patient groups, providing cross-validation in different groups, accounting for differences in populations (Asia and the United States), and multiple equipment manufacturers,” he wrote. “However, the most exciting aspect is the characterization of image texture information.”
Marwick explained that echocardiographic texture, based on brightness and variability, a way of measuring tissue granularity, “has been one of the ‘Holy Grails’ of echocardiography since the dawn of 2-dimensional imaging more than 40 years ago.”
The notion of using radiomic analysis of ultrasound images is exciting, and if realized, it would not only facilitate triage and management on POCUS images, but also add to the size and function measurements of standard echocardiography,” he wrote.
Marwick predicted that as artificial intelligence becomes more integrated into cardiac imaging, these machine-learning models will be able to deliver key information in a matter of minutes.
“The addition of myocardial texture information could potentially improve phenotypic characterization of disease entities, allowing earlier detection of disease, and more specific therapeutic approaches,” he concluded.
Hathaway QA, Yanamala N, Siva NK, et al. Ultrasonic Texture Features for Assessing Cardiac Remodeling and Dysfunction. J Am Coll Cardiol 2022;80:2187–2201.
Marwick TH. Assessment of Myocardial Texture: The Next Frontier in Echocardiographic Quantification. J Am Coll Cardiol 2022;80:2202–2204.
Image Credit: bartekwardziak – stock.adobe.com