Deep learning assesses right ventricle and links findings to cardiovascular outcomes

A deep-learning system called DROID-RV can automatically estimate the size and pumping function of the heart’s right ventricle from routine echocardiogram videos, according to a peer-reviewed study from Mass General Brigham and collaborators. Trained on more than 380,000 video clips from about 17,000 patients, the model estimated right-ventricular diameter to within about 3 millimeters on average and identified reduced right-ventricular function with high accuracy. In a separate linked analysis of primary-care patients, automated predictions of larger or less functional right ventricles were associated with heart failure, atrial fibrillation and mortality.

The right ventricle is the part of the heart that pumps blood to the lungs. Its crescent-like shape and location behind the breastbone make it difficult to assess with cardiac ultrasound, and imaging it takes significant time and expertise. Cardiac MRI is more precise, but it is expensive and not as widely available.

The investigators sought a faster and more consistent way to evaluate right-ventricular size and function from echocardiograms, which are easier to obtain. Mostafa Al-Alusi and Patrick T. Ellinor of the Mass General Brigham Heart and Vascular Institute led the study, published in Nature Cardiovascular Research as “Deep Learning-Enabled Assessment of Right Ventricular Size and Function is Associated with Cardiovascular Outcomes.” The central question was whether a deep learning model could automatically estimate right-ventricular size and detect reduced pumping function directly from routine echocardiogram videos.

The team also investigated whether those automated estimates were associated with a patient’s risk of death, atrial fibrillation or heart failure. To develop DROID-RV, the researchers used a type of model especially well suited to processing image data. They trained it on more than 380,000 echocardiogram video clips from approximately 17,000 patients seen in cardiology clinics.

Working with colleagues from the Broad Institute of MIT and Harvard’s Machine Learning for Health initiative, the team trained the model using both qualitative descriptions of right-ventricular size and function and numerical measurements. The approach was intended to support automated interpretation of right-ventricular information in moving ultrasound images. The researchers tested the model in two groups: about 9,100 primary-care patients from the Mass General Brigham health system and an independent dataset of 831 patients from a hospital in Hungary.

The independent dataset was used to test the model’s image-assessment performance outside the setting in which it had been developed. The outcome analysis was conducted separately in the primary-care cohort, where the researchers linked model predictions with longitudinal health-record data to examine their relationship with later health events. The supplied study summary does not report performance breakdowns by demographic subgroup or a complete assessment of potential biases.

DROID-RV estimated right-ventricular diameter with an average error of about 3 millimeters and identified reduced right-ventricular function with high accuracy, including in the independent Hungarian dataset. In the linked primary-care analysis, predictions of larger right-ventricular size or poorer right-ventricular function were associated with heart failure, atrial fibrillation and mortality. For example, patients with a predicted dilated right ventricle were about twice as likely to die within five years as patients with a predicted normal-sized right ventricle.

These findings describe statistical associations and do not establish that ventricular enlargement caused the outcomes. The outcome associations persisted after the researchers accounted for age, sex, height, weight and left-ventricular function. A further observation was that some DROID-RV predictions were more strongly associated with future cardiovascular disease than measurements made by the echocardiographer who read the original study.

Predicted right-ventricular dysfunction remained independently associated with mortality and atrial fibrillation after the analysis accounted for the echocardiographer’s assessment of right-ventricular function. The investigators said this suggests that the model may contain prognostic information drawn from parts of echocardiogram videos that are not routinely measured. The authors said the model could eventually help reproduce existing echocardiogram measurements and potentially forecast future cardiovascular disease.

They also described possible operational roles for systems such as DROID-RV: providing preliminary measurements, helping echocardiographers read studies faster and flagging abnormal examinations that merit closer review. Another proposed use was to identify patients who might benefit from additional testing with cardiac MRI, especially where MRI capacity or expert readers are scarce or where an immediate assessment is needed in acute settings. These are proposed applications rather than outcomes measured in the study.

The reported evidence does not show that deploying DROID-RV improves clinical decisions, patient outcomes or efficiency in routine care. The investigators tested automated image assessment and examined associations between predictions and health-record outcomes, but they did not report a clinical trial of the tool in use. Although adjustment for age, sex, height, weight and left-ventricular function supported the reported associations, observational data cannot determine causation or establish that the model adds clinically actionable information beyond existing assessment.

Generalizability also remains bounded by the populations and data available in the study. The model was developed using health-system echocardiograms and tested for image assessment in an independent Hungarian dataset, while outcome associations were examined in the primary-care cohort. The supplied evidence does not establish performance in every demographic subgroup, clinical setting or type of patient.

Reproducing established measurements is also distinct from diagnosing disease or selecting treatment: the study evaluated the model’s estimates of right-ventricular size and function and their associations with later outcomes, not its performance as an independent treatment-decision maker. Overall, the study presents DROID-RV as a computational model for extracting right-ventricular information from routine echocardiographic video. Its measured accuracy in estimating diameter and recognizing reduced function was accompanied by outcome associations in the primary-care cohort, where more concerning predictions corresponded with higher subsequent risk.

The observation that predicted dysfunction retained prognostic associations after accounting for the original reader’s assessment hints that the model may capture relevant information beyond conventional measurements. The investigators proposed training future models to predict cardiovascular outcomes directly, rather than only recreate measurements that clinicians already make, but the present study establishes an association and a technical proof of concept rather than a tested clinical benefit.

Subject of Research: Deep-learning assessment of right-ventricular size and function from echocardiogram videos

Article Title: Research Spotlight: AI tool automates echocardiogram assessment of the heart’s right ventricle

Article References: Research Spotlight: AI tool automates echocardiogram assessment of the heart’s right ventricle. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: deep learning, echocardiography, right ventricle, cardiac imaging, heart failure, atrial fibrillation, mortality, cardiovascular risk, clinical decision support, medical artificial intelligence, Research, Spotlight