Using AI to Catch More Cases of Peripheral Artery Disease

by Adrienne Mueller, PhD
October 12, 2022

Peripheral artery disease, or PAD, affects 8-12 million people in the US. The disease is typically caused by plaque buildup on the artery walls of your limbs, which reduces blood flow. PAD is a very common cardiovascular disorder, but it is often underdiagnosed. Indeed, a landmark study showed that as many of 55% of patients could be undiagnosed. This is because the disease can present differently for different patients, and only about 10-20% of patients have ‘typical’ symptoms. Additionally, less than 50% of patients and physicians are even aware of PAD as a disorder.   Failure to diagnose PAD results in poorer outcomes for patients, because treatments aren’t started quickly enough or the condition is misdiagnosed. In terms of patient outcomes, failure to diagnose PAD and begin treatment quickly can result in a higher chance of cardiovascular events like strokes or heart attacks, limb amputation, or mortality. We therefore need better systems to detect and inform doctors about a patient’s risk of having PAD.

Part of a dashboard that displays all patient information, including the AI prediction of PAD risk.

Recently a team of Stanford investigators led by Elsie G. Ross developed an automated tool to identify undiagnosed PAD and, further, tested using a digital dashboard to bring potential PAD cases to doctors’ attention. Their study was recently published in Scientific Reports, co-first-authored by Ilies Ghanzouri, Saeed Amal, PhD, and Vy (Vivian) Ho, MD. The investigators used data already contained in the electronic health records of over 3000 patients with PAD and over 16000 control patients. Electronic health records are a fantastic source of information, because they do not require additional testing or doctors’ visits, and they already contain a large depth and breadth of information about PAD risk factors. Using this rich pool of data, the investigators created a novel deep learning model that successfully identified cases of PAD – and significantly outperformed current alternative models.

In addition, the investigators then developed a dashboard to report patient PAD risk and found that 75% of the physicians they interviewed were open to using a health record-based automated PAD detection system like the dashboard they developed. In fact, the physicians urged the investigators to find ways to take their model even further and also provide risk assessments for clinically complex patients.

In summary, Dr. Ross’s team was able to use machine learning models to take electronic health record data and predict risk of PAD; and, further, to demonstrate that most physicians are interested in using an automated PAD risk detection tool. Once automatic detection of PAD is more adopted in the clinic, future studies will be needed to quantify its impact on patient outcomes.

Additional Stanford Cardiovascular Institute-affiliated investigators who contributed to this study include Lida Safarnejad, John Cabot, Cati G. Brown-Johnson, Nicholas Leeper, Steven Asch, and Nigam H. Shah.

Elsie Ross, MD