AI tools show promise in diagnosing advanced heart failure

Applying artificial intelligence techniques to cardiac ultrasound data may help identify patients with advanced heart failure, a new study has found. The study, led by researchers at Weill Cornell Medical College, Cornell Institute of Technology, Cornell Ann S. Bowers College of Computing and Information Sciences, Columbia University Valléros College of Physicians and Surgeons, and NewYork-Presbyterian, offers the possibility of better care for thousands of patients who are overlooked because their conditions are difficult to diagnose.

Advanced heart failure is currently detected by cardiopulmonary exercise testing (CPET), which requires specialized equipment and trained staff and is typically only available at large medical centers. Partly because of this diagnostic bottleneck, only a few of the estimated 200,000 patients with advanced heart failure in the United States receive appropriate treatment each year. In a new study published March 3 in npj Digital Medicine, researchers tested a new AI-powered method that may remove this bottleneck. The new method uses more easily obtained ultrasound images of the patient’s heart and the patient’s electronic health record to predict the most important CPET measurement, peak oxygen consumption (peak VO2), with high accuracy.

“This opens a promising avenue to more efficiently assess patients with advanced heart failure using data sources already incorporated into routine clinical practice,” said Fei Wang, Ph.D., lead author of the study, associate dean of AI and data science, and the Frances L. Loeb and John L. Loeb Professor of Medical Informatics at Weill Cornell Medical College.

The research was highly collaborative, involving Dr. Wang’s team of informatics and AI experts as well as a group led by Dr. Deborah Estrin, associate dean of Cornell Tech’s School of Impact. And on the clinical side, said Dr. Nir Uriel, director of advanced heart failure and heart transplantation at NewYork-Presbyterian.

Realizing the potential of AI in heart failure treatment

This journal paper is the first published by the Cardiovascular AI Initiative, a broad initiative by Cornell University, Columbia University, and NewYork-Presbyterian to explore the use of AI to improve the diagnosis and management of heart failure. Recent advances in AI have enabled common consumer and business applications, as well as machine learning models trained to detect disease-related patterns in text- and image-based medical data.

“Initially, we assembled a group of more than 40 heart failure experts to tell us where they thought AI would be best applied,” Dr. Uriel said. He is also the Seymour, Paul, and Gloria Milstein Professor of Cardiology at the Columbia University Vagelos College of Physicians and Surgeons, and an adjunct professor in the Greenberg Division of Cardiology at Weill Cornell Medical College.

Using AI on cardiac ultrasound data to help identify patients with advanced heart failure appears to be one of the most promising applications. Dr. Uriel then approached AI experts at Cornell Tech, Cornell Bowers, and Weill Cornell Medicine, and after several years of collaboration, they developed a new machine learning model.

“The close interaction between clinicians and AI researchers on this project has facilitated the development of new AI techniques that would not have been considered otherwise,” said Dr. Estrin, who is the Robert V. Tishman ’37 Professor of Computer Science at Cornell Tech, Cornell Bowers College Professor, and Professor of Population Health Sciences at Weill Cornell University. “So this is a case of medicine shaping the future of AI, and not just AI shaping the future of medicine.”

The AI ​​team led by Dr. Wang, including lead authors Zhe Huang, Ph.D., and Weishen Pan, Ph.D., along with Cornell Bowers students and faculty, developed a multimodal, multi-instance machine learning model that can process several different data types, including regular moving ultrasound images of the heart, associated waveform images showing heart valve dynamics and blood flow, and various items found in electronic health records.

The model was trained on anonymized data from 1,000 heart failure patients seen at NewYork-Presbyterian/Columbia University Irving Medical Center. Once trained, the model was tasked with predicting peak VO2 and effectively determining high-risk status in a new set of 127 heart failure patients from three other NewYork-Presbyterian campuses.

This result was better than any previously reported for AI-based peak VO2 prediction. As a tool aimed at differentiating high-risk patients from other patients, the researchers used a measure that related to the probability that a randomly selected high-risk patient in the sample would have a higher predicted risk than a randomly selected low-risk patient. This figure in this case indicates an overall accuracy of approximately 85%, suggesting that it may be useful in clinical practice.

The research team has already begun planning clinical studies of the new approach for U.S. Food and Drug Administration approval and routine clinical adoption.

“Using this approach to identify many patients with advanced heart failure who would otherwise not be identified would change clinical practice and significantly improve patient outcomes and quality of life,” Dr. Uriel said.