Applying artificial intelligence techniques to cardiac ultrasound data may make it easier to identify patients with advanced heart failure, a new study has found. The study—led by investigators at Weill Cornell Medicine, Cornell Tech, Cornell Ann S. Bowers College of Computing and Information Science, Columbia University Vagelos College of Physicians and Surgeons and NewYork-Presbyterian—offers the prospect of better care for many thousands of patients who may be overlooked due to the difficulty of diagnosing their condition.
Advanced heart failure is currently detected through cardiopulmonary exercise testing (CPET), which requires specialized equipment and trained staff and is typically only available at large medical centers. Due in part to this diagnostic bottleneck, only a few of the estimated 200,000 people in the United States with advanced heart failure get appropriate care each year. In the new study , published March 3 in npj Digital Medicine, the researchers tested a novel AI-powered method that may remove this bottleneck. The new method predicts with high accuracy the most important CPET measure, peak oxygen consumption (peak VO2), using much more easily obtainable ultrasound images of the patient's heart plus the patient's electronic health records.
"This opens up a promising pathway for more efficient assessment of patients with advanced heart failure using data sources that are already embedded in routine care," said study senior author Dr. Fei Wang , the associate dean for AI and data science and the Frances and John L. Loeb Professor of Medical Informatics at Weill Cornell Medicine.
The study was highly collaborative, involving not only Dr. Wang's team of informatics and AI experts but also groups led by Dr. Deborah Estrin , associate dean for impact at Cornell Tech; and on the clinical side, Dr. Nir Uriel , director of advanced heart failure and cardiac transplantation at NewYork-Presbyterian.
Realizing the Promise of AI in Heart Failure Care
The journal paper is the first to emerge from the Cardiovascular AI Initiative , a broad effort from Cornell, Columbia and NewYork-Presbyterian to explore the use of AI to improve heart failure diagnosis and management. Recent advances in AI have enabled not only popular consumer- and business-oriented applications but also machine learning models trained to detect disease-related patterns in textual- and image-based medical data.
"Initially we put together a group of more than 40 heart failure specialists and asked them to tell us where they thought AI could best be applied," said Dr. Uriel, who is also the Seymour, Paul and Gloria Milstein Professor of Cardiology in the Department of Medicine at Columbia University Vagelos College of Physicians and Surgeons and an adjunct professor of medicine in the Greenberg Division of Cardiology at Weill Cornell Medicine.
Using AI on cardiac ultrasound data to help identify advanced heart failure patients seemed one of the most promising applications. Dr. Uriel then approached AI experts at Cornell Tech, Cornell Bowers and Weill Cornell Medicine, who developed the new machine learning model over several years of collaboration.
"The close interaction between clinicians and AI researchers on this project ended up driving the development of new AI techniques that would not have been explored otherwise," said Dr. Estrin, who is the Robert V. Tishman '37 Professor of Computer Science at Cornell Tech, a professor in Cornell Bowers and a professor of population health sciences at Weill Cornell Medicine. "So, this was a case of medicine shaping the future of AI—not just AI shaping the future of medicine."
The AI team led by Dr. Wang, including lead authors Dr. Zhe Huang and Dr. Weishen Pan along with students and faculty at Cornell Bowers, developed a multi-modal, multi-instance machine learning model that can process several distinct data types including ordinary moving ultrasound images of the heart, related waveform imagery displaying heart valve dynamics and blood flow, and various items found in electronic health records.
The model was trained on deidentified data from 1,000 patients with heart failure seen at NewYork-Presbyterian/Columbia University Irving Medical Center. Once trained, the model was then tasked with predicting peak VO2–effectively determining high-risk status—for a new set of 127 patients with heart failure from three other NewYork-Presbyterian campuses.
The results were better than any reported before for AI-based peak VO2 prediction. For tools meant to distinguish high-risk patients from other patients, researchers used a measure that relates to the probability that a randomly chosen high-risk patient in the sample has a higher predicted risk than a randomly chosen lower-risk patient. That figure in this case indicated an overall accuracy of roughly 85%, which suggests it will be useful in clinical settings.
The team has already begun to plan clinical studies of the new approach, which would be needed for U.S. Food and Drug Administration approval and routine clinical adoption.
"If we can use this approach to identify many advanced heart failure patients who would not be identified otherwise, then this will change our clinical practice and significantly improve patient outcomes and quality of life," Dr. Uriel said.
