Ultrasound images of the heart, combined with the power of artificial intelligence, may soon give doctors a far better way to identify patients with advanced heart failure before their condition becomes catastrophic.
A new study published in the journal npj Digital Medicine describes a machine learning model that predicts a key measure of heart failure severity with roughly 85 percent accuracy, using data that hospitals already collect during routine care. For a condition that kills tens of thousands of people every year while going largely undetected, that kind of breakthrough could not come at a more critical time.
The implications are significant. Advanced heart failure is a serious and often fatal condition, yet the vast majority of people living with it never receive the specialized care they need. In the United States alone, an estimated 200,000 people have advanced heart failure, but only a small fraction of them are ever properly diagnosed and treated.
The central reason for this gap comes down to access. The current gold standard for detecting advanced heart failure requires a specialized test called cardiopulmonary exercise testing, which demands expensive equipment, trained staff, and facilities that are typically only available at large medical centers. For patients in smaller communities or under-resourced health systems, that test is simply out of reach, leaving a staggering number of people without a diagnosis that could change the course of their care.
How the ultrasound model works
The research team, drawing on expertise across several leading institutions including Weill Cornell Medicine, Cornell Tech, Columbia University, and NewYork-Presbyterian, developed a model that sidesteps that diagnostic bottleneck entirely. Instead of relying on cardiopulmonary exercise testing, the model analyzes moving ultrasound images of the heart, waveform imagery showing how heart valves and blood are functioning, and relevant data pulled directly from electronic health records.
The specific measurement the model is designed to predict is called peak oxygen consumption. It reflects how efficiently the heart and lungs are delivering oxygen to the body during physical exertion and is considered one of the most reliable indicators of advanced heart failure severity.
By accurately predicting this measurement from routine data sources, the model could allow clinicians at any hospital to flag high-risk patients without needing access to specialized testing. The model was trained using deidentified records from 1,000 heart failure patients and then tested on a separate group of 127 patients from different hospital campuses. Its performance surpassed any previously reported results for AI-based prediction of this particular measurement.
Closing the ultrasound diagnosis gap in heart failure care
What makes this research particularly meaningful is the scale of the problem it is trying to solve. Advanced heart failure patients who go unidentified are not simply missing a diagnosis. They are missing access to treatments, transplant evaluations, and mechanical heart support devices that could extend and dramatically improve their lives. The longer they remain undetected, the worse their outcomes tend to be, and the harder it becomes to reverse the damage already done to the heart and surrounding organs.
The project emerged from a broader effort called the Cardiovascular AI Initiative, which brought together more than 40 heart failure specialists to identify where artificial intelligence could have the greatest impact on patient care. Using AI to interpret cardiac ultrasound data for advanced heart failure detection was considered one of the most promising and practical applications, precisely because it builds on tools and data that already exist in clinical workflows rather than requiring entirely new infrastructure.
The research team has begun planning formal clinical trials, which would be required before the technology could receive regulatory approval and be adopted into routine practice. If those trials confirm what the early results suggest, the model could fundamentally change how advanced heart failure is identified and managed, reaching patients who are currently slipping through the cracks of a diagnostic system that was never designed to find them. For the hundreds of thousands of people living with this condition in silence, that possibility represents something genuinely worth waiting for.
