Artificial Intelligence Can Track Aneurysm, Boost Patient Care

Over 450,000 Americans are currently on dialysis. While this treatment can be lifesaving, frequent needle sticks cause many patients to develop arteriovenous fistula aneurysms, a twisting of the blood vessels. Untreated, these aneurysms can rupture and cause a patient to die. So an artificial intelligence solution that improves diagnostic accuracy could improve patient care.

The later a patient is diagnosed, the more likely he is to experience a negative outcome. So patients need a quick diagnosis and immediate care.

But diagnosing aneurysms is challenging and requires a specialized team — vessel access surgeons.

Not every facility that needs this specialized team has access to these resources. Even those hospitals and clinics that do might have specialists on call, which is less ideal than a regular team being available. When time is of the essence, waiting for providers to arrive on site and make a diagnosis is problematic. 

A simple way for any provider to assess the risk of aneurysm rupture would drastically improve the time it takes for patients to be treated and save lives.

Hanjie Zhang, PhD, and her team at the Renal Research Institute of New York are working on just that.  

“We are trying to develop this automatic identification algorithm so we can do classifications here in the clinic anytime we want,” she told HealthITAnalytics. “Then, if we find something, we can refer the patient to the vessel access surgeon.”

The goal is to refer patients who need surgical intervention sooner, ultimately reducing their risk of rupture.

“We can save a lot of patients’ lives,” Zhang emphasized.

The team first began by developing two algorithms to identify the stage of disease. Because this was a pilot project, they wanted to use two common methods of development for quality assurance.

The first method used an existing, pre-trained algorithm and refined its methods based on the image.

“It’s still customized to you and your unique data set. But it learned from other images,” Zhang pointed out. “This has some benefit especially when you don’t have a large image data set. This is called transfer learning.”

The second method, full training mode, takes a loose algorithm’s structure and completely modifies it based on the data set from which it learns.

“Since it is a pilot project for us, we tried both and both worked pretty well,” revealed Zhang.

The team accurately classified over 90 percent of the images using both methods. And classification took only seconds.

The pilot program relied on a few number of patients. But Zhang is currently working to take the project one step further. She is working to develop an app providers can use in clinic to identify an aneurysm’s stage by simply taking a picture of it.

“The image will automatically go to the cloud and it will trigger the algorithm to make a prediction result that goes back to the app,” Zhang explained. “The clinician is going to see what the stage is and the property of each stage on their screen. It’s going to be very user-friendly. They just need one picture, one click.”

The app will be rolled out at the Renal Research Institute’s clinics. But first, Zhang’s team needs more photos. This will help build the database the algorithm uses to learn and advance, ultimately improving the app’s accuracy.

“We’re going to use this for additional training to refine our models. Once this is all done, we will build the full functional app and go to much larger clinics,” Zhang said.

Images from the clinic will run through the algorithm and the results will be validated by vessel experts for quality assurance.

“Eventually, when this is running in clinics, we’re going to have people frequently check how the model performs to make sure it really works well,” said Zhang. “You could do a random checking to see if it works well or you could download the results and compare.”

Throughout development and roll out, patient privacy will remain a priority. The methods and algorithm were approved by the institute’s chief security officer, Zhang noted, and the app will be HIPAA compliant.

The app can reduce provider burden while improving diagnostic time and accuracy. So Zhang hopes the tool will eventually be used by even larger clinics across the country to make complex diagnostic decisions in seconds thanks to artificial intelligence.