Alphabet Inc.’s Google has developed artificial intelligence that can detect a condition that causes blindness in diabetic patients, though tests in India demonstrate the challenges of transferring such technology from the lab to the doctor’s office.
The tech giant’s Automated Retinal Disease Assessment tool is trained to diagnose diabetic retinopathy, a complication of diabetes, by reading images of patients’ retinas. Studies show that the artificial intelligence can be as effective as a doctor at identifying symptoms.
Google is testing ARDA in India where at least 60 million people have diabetes and often don’t know they should be screened for the eye condition. A shortage of ophthalmologists means reaching patients is difficult, and doctors say the technology could help to fill the gap.
Google is the latest tech firm to explore using AI in health care. Amazon.com Inc., International Business Machines Corp. and UnitedHealth Group Inc. sell software that mines patients’ records, and IBM developed AI that detects melanoma on images of skin.
By using technology similar to algorithms that recognize people or dogs in photos, ARDA detects signs of the illness–lesions, burst blood vessels and patches of yellow on the retina–and grades their severity.
“This is a technology that can make a really big difference,” said Lily Peng, a Google product manager and one of the leaders of the team developing ARDA at the company’s California offices.
While ARDA is effective working with sample data, according to three studies including one published in the Journal of the American Medical Association, a recent visit to a hospital in India where it is being tested showed it can struggle with images taken in field clinics. Often they are of such poor quality that the Google tool stops short of producing a diagnosis–an obstacle that ARDA researchers are trying to overcome.
The stakes are high. If diabetic retinopathy is caught early it can be kept at bay through monitoring and management of the diabetes, said R. Kim, an Indian ophthalmologist who runs the Aravind Eye Hospital in Madurai, Tamil Nadu, where Google is testing ARDA. More advanced stages need laser surgery that can stop progression. If it isn’t treated, the condition can cause blindness.
For Google, diabetic retinopathy was an ideal case study because experts agree on its symptoms. With skyrocketing rates of diabetes, India provided the right testing ground. As incomes have grown, so have citizens’ waistlines. The prevalence of the condition in India increased 64% from 1990 to 2016, according to government data.
“I am not proud to say, but India is becoming the diabetic capital of the world,” Dr. Kim said.
Yet the country has just 11 ophthalmologists for every million people, compared with 59 in the U.S., according to the International Council of Ophthalmologists. Training the number needed is “a race you can’t win,” he said.
In many cases, diabetic patients in India visit a doctor only when they have already lost some vision, Dr. Kim said. He hopes ARDA will reach more patients through doctors’ offices or pharmacies, pre-empting the need for his ophthalmologists to run screening camps in rural areas.
To create ARDA, Ms. Peng’s team obtained more than 1.6 million images of retinas that were each evaluated by ophthalmologists. Google used those graded images to train the algorithm. Now the team is grappling with how to make it work with low-quality images produced by the sort of equipment affordable in developing countries such as India.
On a November day in a telemedicine center at Dr. Kim’s hospital, ophthalmologists wearing headsets waited for 73 local eye clinics to connect patients in rural areas via webcam.
Ophthalmologist Anusha Arunachalan got a call from a clinic an hour from Madurai. A 39-year-old diabetic homemaker was having trouble seeing.
Six images sent by a clinic technician showed the patient’s retinas covered by pools of blood. Dr. Arunachalan diagnosed proliferative diabetic retinopathy–an advanced stage of the condition. Then Google gave it a try. It graded two of the six images as proliferative–though one was deemed of poor quality–and two others as severe. The tool was unable to grade two of the photos accurately because they were unclear.
While Dr. Arunchalan said it also was hard for her to read some of the images, “One [clear] image is enough for me. I don’t need three images” for each eye to make a diagnosis, she said.
Doctors at the telemedicine center said they often look at the clear parts of faulty images to make a diagnosis, while the algorithm sometimes can’t.
It isn’t a unique problem, said John Quackenbush, professor of computational biology at the Harvard T.H. Chan School of Public Health. Algorithms can be trained on sample images but real-world photos can be different.
“I don’t think it is insurmountable,” he said, on hearing about Google’s challenges.
Google is also testing the algorithm with another partner in India and one in Thailand and is working on making ARDA work with different-quality images, Ms. Peng said. But the challenge is also ethical.
If Google allowed the algorithm to make a diagnosis from blurred images, it could miss small lesions that appear in the early stages of the condition, she said. Google must decide how bad an image can be before ARDA refuses to grade it.
“It’s a trade-off. We want them to be able to use cameras that are a little harder to use but at some point it should move into something where it is ungradable,” Ms. Peng said.
Nonetheless, when ARDA gets a good image, it can identify early signs of the condition that doctors may have missed, Dr. Kim said.
Whereas ARDA does it in seconds, a human can take three minutes to grade an image plus the time taken to open it on a computer, Dr. Kim said. A busy doctor, he said, could very well miss tiny red lesions in a patient’s eye.
