YuLin Zhen, Staff Photographer

A team of researchers at the CarDS — Cardiovascular Data Science — Lab at the School of Medicine published a study detailing their new AI tool that can screen for cardiomyopathies years before typical diagnosis by medical professionals. 

Founded in 2020, the CarDS Lab has been at the forefront of creating AI-based applications to improve medical diagnoses of heart muscle diseases for the past five years.

“We want to design AI tools that we can use with tests that are easy to perform and that can be easily available in the community,” Dr. Evangelos Oikonomou told the News. “We don’t necessarily want to build AI tools for technologies that might be very hard to come across or find and are only restricted to very specific, high resource settings.”

The team of researchers began working on the grant proposal almost two years ago and spent around a year and a half compiling data from different hospitals to train the AI model. 

With an interest in developing AI models that rely on data provided by accessible diagnostic tools, the team focused on point-of-care ultrasound — a portable ultrasound test that can be performed by plugging an ultrasound rod into a smartphone to obtain a detailed image of the heart. The test is widely available to medical providers but is mainly used for crude assessments of how the heart works and if there are any obvious abnormalities.

According to Oikonomou, even with the use of point-of-care ultrasound, many abnormal heart conditions go undetected. Recognizing patterns that indicate cardiomyopathies requires extensive training and expertise, and it is difficult for expert operators to carefully examine every ultrasound image. 

“We see that the model is actually able to pick up on patterns that are visible to the human eye, but probably not detectable by an untrained operator,” Oikonomou said. “But it also goes a bit beyond that. It even seems to detect those conditions way before clinicians actually suspect what’s going on.”

The study focused on two main forms of cardiomyopathies — hypertrophic cardiomyopathy and amyloid cardiomyopathy, both diseases that make it difficult for the heart to pump blood. 

Oikonomou noted that medical professionals have realized that the condition is much more common than previously thought and that the lack of expert operators and effective diagnosis tools made the diseases difficult to detect in the past. 

To train their AI model, the team fed real-world data collected from more than 30,000 patients across the Yale Health System over a decade.

After the team created a functional training model, they started testing videos from patients who were not seen by the AI model beforehand and who were screened with point-of-care ultrasound devices across the emergency rooms of the Yale Health System, the Yale New Haven Health System and the Mount Sinai Health System.

While hypertrophic cardiomyopathy is inborn and amyloid cardiomyopathy is acquired during one’s lifetime, both diseases are progressive, meaning that the model can actually pick up on earlier stages of the disease that cannot be detected by the human or untrained eye.

“We found that the algorithm could pick up the disease at an average of two years before the disease was eventually diagnosed in real-world practice,” said Oikonomou. “We also found that there were a lot of patients that were never actually tested for any of those conditions but were flagged as high risk by our models, and these patients went on to have worse outcomes.”

Oikonomou noted that the model stratifies risk in patients, with those flagged positively being much more likely to have the disease progress faster than those flagged negatively. 

AI models accurately detected two types of cardiomyopathy with 0.95 and 0.98 AUROC, a performance metric used to evaluate binary classification models where achieving a 1 is considered a “perfect model,” and were received well in testing at Mount Sinai Hospital System located in New York.

The CarDS lab published this AI application for anyone to access for research purposes. 

“We need to make use of AI. We need to leverage AI to make those technologies more accessible, more scalable,” Oikonomou told the News.

AI has become increasingly integrated into medical practices. However, the successful widespread implementation of such technologies depends on their affordability and availability.

The CarDS lab used point-of-care ultrasound because it is easily applicable for community-based screenings of abnormal heart disorders, costs less than $2,000 and can easily be plugged into a smartphone.   

Oikonomou emphasized the importance of the team’s work with external collaborators at Mount Sinai, who independently ran the AI model without any transfer of their data. 

“This multi-site validation confirms that the model will work reliably when exposed to new settings and patient distributions beyond what was seen during model development,” Gregory Holste, a doctoral student in the CarDS Lab, said.

Oikonomou shared that they are in the process of designing a clinical trial where some providers are going to have access to these AI tools and others will not. Over the years, the results from this trial should give insight into the clinical value of using AI technologies to detect cases of abnormal heart muscle disorders.

More information regarding the CarDS Laboratory’s research can be found here.