Machine Learning Study Identifies Hidden Risk Groups in Aortic Regurgitation, Overlooked Danger for Women

A new multicenter study led by Houston Methodist and Rice University researchers has identified four distinct patient profiles in aortic regurgitation — including a high-risk group of women whose disease often appears milder by traditional measures but is linked to significantly worse outcomes.

The findings, published in JACC: Cardiovascular Imaging, suggest that long-standing "one-size-fits-all" approaches to managing the valve disorder may miss key differences that matter for survival.

Enhancing risk stratification

Aortic regurgitation, or AR, occurs when the aortic valve fails to close tightly, allowing blood to leak backward into the heart. Over time, the condition can enlarge and weaken the left ventricle, but patients progress at widely different rates — a clinical pattern that cardiologists have long observed but previously could not predict.

"This is a question we come across constantly," said Dr. Maan Malahfji, a cardiologist at Houston Methodist and first author on the study. "Some patients tolerate significant AR for years without symptoms, while others develop problems much earlier. Yet guidelines still trigger surgery at the same thresholds for everyone."

To better capture these differences, the team applied an unsupervised machine learning algorithm to clinical and cardiac MRI data from 972 patients across four U.S. centers. The approach allowed the computer to group patients by similarity — without being told who lived or died — and then compare outcomes across clusters.

Rice's statistical team led the development and validation of the clustering pipeline, which handled missing data and complex variable interactions more effectively than traditional models.

The algorithm consistently revealed four phenoclusters, each with unique clinical features and survival patterns.

One cluster included younger, mostly male patients — often with bicuspid valves — who showed extensive ventricular enlargement but excellent outcomes. Another included older men with multiple comorbidities and high mortality. A third cluster captured patients with advanced scarring and dysfunction, a group that predictably fared poorly.

Surprise findings for women

But the most striking finding was a female-predominant phenotype. These patients showed less dramatic heart enlargement and lower measured AR severity yet had 20% mortality, similar to the sickest cluster of patients.

"For women, the ventricle often doesn't dilate as much," Dr. Malahfji said. "Their hearts are smaller to begin with, and indexing for body size is still an imperfect correction. What we see clinically — and the model confirms — is that women may develop shortness of breath or fatigue earlier, but their imaging doesn't always reflect that severity. That can lead to under-referral for surgery."

The researchers also created a prototype risk calculator to help clinicians identify which phenocluster a patient fits into — a tool that could eventually inform timing of intervention. While the calculator doesn't replace existing guidelines, Dr. Malahfji said even incremental improvements in prediction can impact care.

"It gives clinicians a clearer sense of long-term trajectory," explained Dr. Malahfji. "As less invasive valve therapies expand, thresholds for intervention may shift, and better phenotyping helps ensure that high-risk patients, especially women, aren't overlooked."

The project was enabled by the Rice–Houston Methodist Digital Health Institute, which aims to accelerate AI-driven advances in clinical care.

"The takeaway here is positive," concluded Dr. Malahfji. "With more precise risk stratification, we can personalize surveillance and intervene earlier for those who need it, while sparing low-risk patients unnecessary procedures."