AI model reads heart ultrasounds to size up the right ventricle

On a routine heart ultrasound, echocardiographers have to pick out a chamber that is hard to see. The right ventricle, the part of the heart that pumps blood to the lungs, has a crescent shape and sits behind the breastbone, so measuring it takes significant time and expertise. The more precise alternative, cardiac MRI (magnetic resonance imaging), is expensive and not as widely available. A team at the Mass General Brigham Heart and Vascular Institute set out to automate that measurement.

The tool, DROID-RV, is a deep learning model, a type of AI especially good at processing images. Its authors trained it on more than 380,000 echocardiogram video clips from about 17,000 patients seen in cardiology clinics. With colleagues from the Machine Learning for Health initiative at the Broad Institute of MIT and Harvard, they taught it to learn from both written descriptions of the chamber and numerical measurements. They then tested it on about 9,100 primary care patients from their own health system, and on an independent set of 831 patients from a hospital in Hungary. Finally, they linked its predictions to the patients' health records followed over time.

On measurement, DROID-RV estimated right ventricular diameter to within about 3 millimeters on average, and it identified reduced pumping function with high accuracy, including on the outside data set. The predictions also tracked later illness. Patients predicted to have dilated right ventricles were twice as likely to die within five years as patients predicted to have normal size. Larger size or worse function was associated with heart failure, atrial fibrillation (an irregular heartbeat) and death. Those links held after accounting for age, sex, height, weight and left ventricular function, the main pumping chamber.

The most striking result, the authors say, is that the predictions were sometimes more strongly tied to future disease than the measurements the echocardiographer had made on the same scan. Predicted dysfunction stayed linked to death and atrial fibrillation even after that assessment was taken into account. The authors suggest the model may draw on parts of the video that are not routinely measured. These are associations observed in patient records; the study does not show that acting on the predictions changes outcomes.

For echocardiographers, the nearest gain is speed: the tool could provide preliminary measurements and flag abnormal exams that need a closer look. It could also point patients toward a costlier MRI when one is needed, which matters most where MRI machines or expert readers are scarce, and in acute cases that call for an immediate assessment. The authors say tools like DROID-RV could eventually also predict future cardiovascular disease, and that future models could be trained on outcomes directly. They give no date for any of this, and DROID-RV remains a research result.