Using Azure Machine Learning to Combat Preventable Blindness

Diabetes is the leading cause of preventable blindness in the United States, but there was no easy way to diagnose diabetic vision damage through primary care providers. That’s why IRIS used Microsoft Azure to help create a platform that can identify diabetic retinopathy before patients suffer from vision loss. Using Azure Machine Learning Package for Computer Vision, the IRIS platform processes images quickly and accurately so doctors can share data with patients and other clinicians, better prevent diabetic blindness, and help reduce healthcare costs.

“We went from zero to 300,000 patients examined in under five years—there is no way we could have done that without Azure.” Jonathan Stevenson: Chief Strategy and Information Officer, IRIS

Combating a devastating disease

While most people with diabetes can control their disease with diet and therapy, they can still also suffer from a range of serious complications. Because diabetes damages blood vessels throughout the body, people with diabetes have a higher risk of developing diabetic retinopathy, caused by damage to the blood vessels in the back of the eye. In fact, diabetes is the number one cause of preventable blindness in the United States.

More than 29 million people have diabetes in the United States—and that number is growing. Most of them probably get their eyesight checked regularly, but diabetic retinopathy can advance significantly before it begins to impact vision. A patient can have twenty-twenty vision and still be on the verge of losing their eyesight due to blood vessel damage or aneurysms. This state of risk can be identified with a simple exam, performed in a number of settings, with a number of different tools. However, primary care providers (PCPs) traditionally do not perform this type of noninvasive exam during a clinic visit. Typically, only ophthalmologists and some optometrists have the technology to test for diabetic retinopathy. But why would patients schedule an eye appointment if they were asymptomatic?

In 2011, retinal surgeon Dr. Sunil Gupta gathered a team to meet these challenges and try to end diabetic blindness. They developed the Intelligent Retinal Imaging Systems (IRIS) solution and launched a business by the same name. IRIS captures a high-resolution image of the back of the patient’s eye that clinicians use to identify diabetic retinopathy and determine if the patient is at risk of vision damage.

Finding a scalable cloud solution in Azure

To reach every patient with diabetes in the United States, IRIS needed a solution that was fast, scalable, and easy to deploy, so it built the solution as a cloud-native service on Microsoft Azure. Now, patients can get tested for diabetic retinopathy through their PCP during a regular visit.

A clinician, or any trained clinic staff member, uses the IRIS system to take an image of the retina (the back of the eye). The image is sent to Azure Service Bus where it is processed for feature detection and image enhancement, such as morphing the colors to see pathology. The image data continues to Azure Machine Learning Package for Computer Vision, which uses deep learning algorithms to identify and categorize any pathology in that image. IRIS uses NVIDIA graphics processing unit (GPU) computing to run Machine Learning Package for Computer Vision. Then IRIS uses services built with Azure Functions to take the image from Service Bus and transfer the data into Azure SQL Database, which triggers a message to the clinician that a result is available and ready to be shared with the patient.

“Using Azure Machine Learning Package for Computer Vision, we were able to reach 97 percent accuracy on our production images,” says Jocelyn Desbiens, Data Scientist at IRIS. “And deploying the models in Azure is incredibly fast and easy.”

Using Azure services has proved to be as scalable as IRIS hoped. “We went from zero to 300,000 patients examined in under five years—there is no way we could have done that without Azure,” says Jonathan Stevenson, Chief Strategy and Information Officer at IRIS. “Utilizing Azure Machine Learning Package for Computer Vision powered by NVIDIA GPUs to run our experiments allowed us to get to success much faster than we could have with traditional platforms.”

Successfully deploying at MyHealth First Network

MyHealth First Network (MyHFN) is a large, clinically integrated network of providers based in South Carolina with a mission to transform care, promote value, and enrich the patient experience. Beginning in early 2018, MyHealth First Network started using IRIS at four of its Greenville Health System practices. IRIS staff worked onsite during the first week to train clinicians how to use the solution.

Whenever diabetic MyHFN patients visit their PCP, an IRIS retinal exam is recommended. The exam itself takes less than five minutes and IRIS guarantees less than 24 hours to process the results, but often delivers them early. “Thanks to Azure Machine Learning, our PCPs get detailed results back quickly,” says Dr. Pat Marshall, President at MyHealth First Network. “Not infrequently, we receive the results before the conclusion of the visit, which facilitates a timely discussion with the patient.”

Slowing and preventing blindness

If patients are diagnosed early enough, it’s possible to significantly slow or prevent blindness with treatment. In the first three months using IRIS, MyHealth First Network examined 665 patients. Of those, 139 were identified with pathology, and 39 of those had sight-threatening disease. “With IRIS, we can identify the problem early and prevent about 90 percent of the loss of vision,” says Dr. Marshall. “Remember, these patients were unaware of this problem when they walked into their PCP’s office. The impact is impressive.”

MyHealth First Network has also seen IRIS exam results influence patients’ lifestyles. For example, using IRIS, a PCP diagnosed a patient with mild diabetic retinopathy and showed her the picture of her retina—where she could see actual aneurysms. “The clarity of the image from IRIS impacted the patient so much that she completely changed her lifestyle, exercising more and eating more appropriately for her diabetes,” says Stevenson. When they examined her six months later, she had no new aneurysms, and the old ones were healing.

Reducing healthcare costs

When any patient is diagnosed early with just a few small aneurysms, the treatment costs $200 to $400. If the patient is diagnosed with severe pathology, the class of drug needed is far more expensive, ranging from $1,500 to $3,000 per treatment.

“Patients who reach the cusp of vision impairment face a huge threat to their quality of life,” says Stevenson. “It also costs the healthcare system almost 10 times more to make sure that patients maintain their vision—costs we’re minimizing with our solution in Azure.”

On the provider side, IRIS connects to electronic medical records so that clinicians can easily receive results from an IRIS scan. They simply get a notification and then can make a referral if needed. “Doctors are very busy, and it can be hard to get buy-in from them on new technology because they are concerned it will take more time and money,” says Dr. Marshall. “By using Microsoft technology, IRIS has created a seamless and cost-effective solution that clinicians embrace to dramatically improve outcomes for patients with diabetes.”

Find out more about IRIS on Twitter, Facebook, and LinkedIn.

“By using Microsoft technology, IRIS has created a seamless and cost-effective solution that clinicians embrace to dramatically improve outcomes for patients with diabetes.” Dr. Pat Marshall: President, MyHealth First Network