Machine-learning to predict right ventricular heart failure
in LVAD patients
by Roxanna Van Norman
December 2, 2021
A Stanford study found using a deep learning system could predict right ventricular failure after cardiac surgery, significantly outperforming a team of human experts conducting the same evaluation.
A study led by researchers at the Stanford Department of Cardiothoracic Surgery found that analyzing real-time videos of the heart with artificial intelligence could predict right ventricular (RV) failure after heart surgery more accurately than a team of human experts conducting the same evaluation.
"We took on a challenging clinical problem of predicting who was likely to get post-operative right heart failure after implantation of a left ventricular assist device [in a patient]," said Rohan Shad, MD, a postdoctoral research fellow in the Department of Cardiothoracic Surgery and the lead author of the study.
The researchers programmed a deep learning system to analyze ultrasounds of the heart taken before surgery and benchmarked the system against board-certified
clinicians using contemporary clinical risk scores. From this, they identified which patients under consideration for left ventricular assist device (LVAD) would likely experience RV failure after surgery.
A challenging clinical problem
While a heart transplant remains the gold standard for treating patients with end-stage heart failure, an alternative option is receiving an LVAD, a battery-powered mechanical pump implanted in the patient. Unfortunately, a third of all patients implanted with LVADs develop a clinically significant degree of RV failure soon after the procedure.
"We wanted to know if there is a way to figure out, before we even take anybody into the operating room, which patients are more likely to suffer from these right-sided heart problems," said William Hiesinger, MD, Assistant Professor of Cardiothoracic Surgery and Surgical Director of the Stanford Mechanical Circulatory Support Program, and senior author on the study. "If you can, then you might be able to provide interventional therapies quicker and faster, or you might be able to do something before you take them into the operating room to improve outcomes down the road."
By training their artificial intelligence (AI) system to identify abnormal motions in the heart videos, Dr. Shad explained, it could conclude from the extensive, unstructured video data who was more likely to suffer from right heart failure after surgery.
The research team looked at various clinical scoring systems, including two contemporary risk scores, to identify patients at risk for RV failure in LVAD candidates. The team then compared the performance of their AI system against clinical risk scores.
Artificial intelligence and clinical outcomes
The study included a dataset of 941 patients who had LVAD implants from three medical centers: Stanford Health Care, Spectrum Health in Minnesota, and the Houston Methodist Hospital in Texas. The dataset was separated into two groups – about one-third (182 patients) were identified with right ventricular heart failure and the remaining two-third (541 patients) were otherwise normal.
The researchers looked at the area under curve (AUC) of the receiver operator characteristic curves, a metric typically used for evaluating machine learning models. An AUC of 0.5 was equivalent to a 50-50 chance of a correct prediction, and an AUC of 1.0 meant the evaluation was closer to perfectly predicting a patient having an RV failure every time.
Most clinical risk scores predicted the onset of the right heart failure problem with an accuracy of about coin flip. The study used CRITT and Penn Scores, which calculated AUCs of 0.616 and 0.605, respectively. For the newly developed deep learning system, the AUC was 0.729 – closer to providing a more accurate prediction of right heart failure.
The team then compared the predictive performance of their AI system against a group of cardiologists who were blinded to the outcomes of these patients. Using their observations of imaging data presented before them, the cardiology team attempted to predict which patients were likely to develop RV failure. The results from the clinicians’ evaluations showed AUCs ranging between 0.525–0.571.
"It is possible that our AI system was able to see something that humans weren't good at characterizing in a concrete fashion for this clinical problem," said Dr. Shad.
Better outcomes for LVAD patients
Predicting which patients will develop RV failure after the implantation of an LVAD is only one such challenge that Dr. Hiesinger and his team are attempting to study with these novel techniques.
"For cardiac surgery, in general, it's an exciting new way of looking at clinical problems that nobody has really ever done before," said Dr. Hiesinger. Although they were working with a small but diverse dataset across multiple centers, the team was surprised by the significant difference in the results. They aim to expand the sample pool across multiple centers and broaden the scope to help clinicians make surgical decisions.
"What excites me the most is thinking about where we can go from here and how we can design deep learning systems that can better represent these complex diseases," said Dr. Shad.
The study's authors included Nicolas Quach, Robyn Fong, Patpilai Kasinpila, MD, Cayley Bowles, MD, Miguel Castro, Ashrith Guha, Erik E. Suarez, Stefan Jovinge, Sangjin Lee, Theodore Boeve, Myriam Amsallem, Xiu Tang, Francois Haddad, Yasuhiro Shudo, MD, PhD, Y. Joseph Woo, MD, Jeffrey Teuteberg, MD, John P. Cunningham, and Curtis P. Langlotz, MD, PhD.
This project was supported by a Stanford Artificial Intelligence in Medicine and Imaging (AIMI) Seed Grant and Cloud Compute Credits from Google Cloud. Dr. Shad was supported in part by the American Heart Association Postdoctoral Fellowship Award.
The findings from the study were published on August 31 in Nature Communications.