Deep learning model automates right ventricular assessment from echocardiograms

Mostafa Al-Alusi, MD, MS, and Patrick T. Ellinor, MD, PhD, of the Mass General Brigham Heart and Vascular Institute, are the lead and senior authors of a paper published in Nature Cardiovascular Research, "Deep Learning-Enabled Assessment of Right Ventricular Size and Function is Associated with Cardiovascular Outcomes."

Q: What challenges or unmet needs make this study important?

The right ventricle (RV) - the part of the heart that pumps blood to the lungs - is hard to assess with cardiac ultrasound because of its crescent shape and position behind the breastbone. For this reason, imaging the RV takes significant time and expertise. More advanced imaging such as cardiac MRI is more precise, but it is expensive and not as widely available.

We wanted a fast, consistent way to evaluate the right ventricle using echocardiograms, which are easier to obtain.

Q: What central question(s) were you investigating?

In our study, we asked, "Can a deep learning model automatically estimate the size of the right ventricle and detect reduced pumping function directly from routine echocardiogram videos?"

We also wanted to explore whether those automated estimates could tell us anything about a patient's risk of death or risk of developing atrial fibrillation or heart failure, conditions where RV size and function are important to prognosis.

Q: What methods or approach did you use?

Using a type of model that is especially good at processing image data, we developed DROID-RV and trained it on more than 380,000 echocardiogram video clips from about 17,000 patients seen in cardiology clinics.

In collaboration with colleagues from the Machine Learning for Health initiative at the Broad Institute of MIT and Harvard, we trained the model to learn from both qualitative descriptions of right ventricular size and function, and also from numerical measurements.

We tested it in about 9,100 primary care patients from our health system and in an independent dataset of 831 patients from a hospital in Hungary. We then linked our model's predictions with longitudinal health record data from the primary care patients to see whether our model's predictions were linked to later health outcomes.

Q: What did you find?

DROID-RV estimated right ventricular diameter to within about 3 millimeters on average and identified reduced right ventricular function with high accuracy, including in the outside dataset.

Model predictions of larger RV size or worse RV function were associated with heart failure, atrial fibrillation and mortality. For example, patients with predicted dilated right ventricles were twice as likely to die within five years, compared to patients with predicted normal RV size. The associations between model predictions and future cardiovascular disease persisted even after accounting for factors such as age, sex, height, weight and left ventricular function.

Q: What are the real-world implications, particularly for patients?

Tools like DROID-RV can help echocardiographers read echocardiograms faster by providing preliminary measurements and flagging abnormal exams that need a closer look. They could also help identify patients who may benefit from additional testing with more expensive MRI scans to better characterize their RV, especially in situations where MRI machines or expert readers are scarce or in acute settings where immediate assessment is needed.

Eventually we may be able to use tools like DROID-RV to not only recreate echocardiogram measurements but also to better predict future cardiovascular disease.

Q: Were you surprised by any of the findings?

One of the most interesting findings was that our model's predictions of RV size and function were sometimes more strongly associated with future cardiovascular disease than the measurements made by the echocardiographer who read the original study. For example, predicted RV dysfunction remained independently associated with mortality and atrial fibrillation even after accounting for the echocardiographer's own assessment of RV function. This is exciting because it suggests the model's predictions contain information relevant to prognosis beyond what it was explicitly trained to predict, possibly drawn from parts of the echocardiogram video that are not routinely measured.

It also raises the possibility of training future models to predict cardiovascular outcomes directly, rather than only recreating measurements that clinicians already take from the echocardiogram.