Using Artificial Intelligence to Optimize Aortic Stenosis Echocardiographic Surveillance

By Michael H. Crawford, MD, Editor

SYNOPSIS: A machine learning model to predict the frequency of echocardiographic exams in patients with mild-to-moderate aortic stenosis necessary to accurately detect the onset of severe stenosis performed with a high degree of accuracy would reduce the number of unnecessary echocardiograms specified by current guidelines.

SOURCE: Sánchez-Puente A, Dorado-Díaz PI, Sampedro-Gómez J, et al. Machine learning to optimize the echocardiographic follow-up of aortic stenosis. JACC Cardiovasc Imaging 2023;16:733-744.

Since the rate of hemodynamic progression of aortic stenosis (AS) is heterogeneous and the onset of symptoms variable, periodic echocardiographic surveillance is necessary, but the optimal timing of these studies is controversial. These investigators from Spain sought to develop three machine learning models to determine the timing of repeat echocardiographic studies to optimally detect the onset of severe AS.

Researchers analyzed the echocardiographic and electronic medical record databases of the University Hospital of Salamanca from 2007 to 2019 for patients with mild-to-moderate AS and who had undergone at least two subsequent echoes. The authors used a transaortic valve gradient of > 4.0 m/s to define severe AS. For the machine learning program, they used 92 clinical and echocardiographic variables.

After analyzing the Salamanca data, investigators repeated the analysis using similar data from a second university hospital in Madrid. To predict the development of severe AS, researchers defined premature as severe AS developing longer than six months after predicted, timely as within six months, and untimely as after six months.

The mean age of the 1,638 patients entered in the study was 73 years, 52% were men, 69% showed mild AS on entry, and 31% showed moderate AS on entry. The patients averaged 1.8 echoes over a mean follow-up interval of 3.6 years. Severe AS developed in 19% of patients (40% of these at year 1, 33% at year 2, and 27% at year 3). The receiver operating curve (ROC) for the prediction of severe AS at year 1 was 0.9 and 0.92 for years 2 and 3.

In the external application data set from Madrid, the corresponding ROCs all were 0.85. With the machine learning system, 29% of patients underwent a timely follow-up echo, 4% an untimely echo, and 1% a premature echo per patient. The performance of the model in the specified subgroups of age younger than 55 years (likely bicuspid valves) and low-flow low gradient AS showed excellent ROCs (0.84-1.0).

A simulated application of the model in the Madrid cohort showed a 49% reduction in echoes using the European guidelines and 13% reduction using the American guidelines. The authors concluded applying a machine learning program in patients with mild-to-moderate AS provides a personalized timing schedule for follow-up echoes to accurately detect the development of severe AS. This could reduce the number of unnecessary echoes specified by the European and American guidelines.

COMMENTARY

It seems the news lately has been filled with articles about artificial intelligence (AI) and its various applications, with commentary by scientists and lay gurus like Elon Musk. Many of these articles highlight the potential of AI for harm as well as good. This paper from Spain falls in the latter category — and is quite relevant, given the growing number of AS patients as our populations age.

Many of us have realized that our guidelines, especially the European ones (every year for mild to moderate AS and every six months for asymptomatic severe AS) result in a lot of unnecessary echoes for patients we are following for the development of severe AS. After all, the guidelines in this area are based on expert opinion (class C) not class A randomized trial data. We also know the hemodynamic progression of AS is heterogenous and the onset of symptoms variable. Thus, some form of follow-up imaging scheme is necessary. The guidelines are just that, and many of us use our “gut” or experience to tell us what follow-up imaging period is appropriate for each patient. However, an AI approach might integrate more complex data than our brains and create a better schedule of follow-up echoes that maximally eliminate unnecessary echoes without compromising patient care. This study from Spain demonstrates the high degree of accuracy AI can accomplish for predicting when to order follow-up echoes to precisely detect the onset of severe AS.

Still, there were limitations to the Spanish study. The authors used only peak velocity to diagnose severe AS. Researchers could have integrated velocity, pressure gradient, and calculated valve area to determine severity, given the challenges of obtaining accurate Doppler data in some patients. However, the investigators did study subgroups of their population using mean gradient and the dimensionless index instead of peak velocity. The authors did not find any substantial difference in the predictions. The authors also did not consider the degree of calcification, which can be an accurate predictor of severity. Finally, researchers confined their model to echo data, even though cardiac CT and MR are gaining ground for assessing AS severity.

AI is the wave of the future in medicine, but we must be mindful of the timely adage: garbage in, garbage out. These programs depend on the quality of the clinical data. If the information is inadequate, there is a possibility of harm.