Diabetes-induced blindness is a rising issue, according to a recent article by UW School of Medicine assistant professor Aaron Lee. Due to an increased demand for care and few ophthalmologists and optometrists who can identify the condition, scientists are looking for ways to effectively automate screenings using artificial intelligence to maximize coverage.
According to Lee, one of the factors making diabetic retinopathy such an important issue is that it is hard to detect, and as a result, may be identified late, causing irreversible damage.
In an ideal world, everyone would go in for routine screenings, and retinopathy would not be an issue. But, Lee explained, it isn’t that simple.
“The screening rate of people who need an eye exam is very low in the United States,” Lee said. “And most people don’t know, and those who do often forget.”
For those who do get an exam, the wait time between the exam and the decision on whether they need further examination and treatment by a professional can be subject to long delays, deprioritization by patients, or patients simply ignoring the referral.
The chain of events that happens in between yearly eye exams, in conjunction with the low supply of qualified professionals, has increasingly pushed researchers to look at machine learning approaches to diabetic retinopathy screening, which would increase the amount of patients who do get timely treatment.
In the study, Lee and his co-authors examined seven different image classification algorithms on 311,604 images from 23,724 different patients with diabetes.
One of the challenges to finding an algorithm that is appropriate for the context is that researchers have to balance the negative predictive capability of an algorithm with its sensitivity, Lee explained.
“We chose negative predictive value for this study because it is a screening problem,” Lee said. “We wanted to be sure that when the algorithm says that there is no disease, there is truly no disease.”
But one of the challenges for researchers is finding the right negative predictive value and sensitivity of an algorithm without making the algorithm useless. If both are set too high, the algorithm will refer everyone it screens for a test.
For Lee and his team, the results of the study found varying rates of sensitivity, from 50.98% to 85.90%, but despite the algorithms performing differently, Lee stressed that they all performed better than a human could, highlighting the importance and potential effectiveness of these devices for real-world use.
Of the algorithms that performed poorly, Lee and other researchers are studying what could be improved to make them viable for clinical use. Lee explained that what researchers are looking for is algorithms that have appropriate predictive levels, while also being durable and scalable for real-world settings.
Currently, there are two FDA-approved algorithms that can be used in clinical settings. The process for getting an algorithm approved by the FDA is long and rigorous, according to Lee, but there is hope that more algorithms like these will be operationally ready in the future.
Reach reporter Thelonious Goerz at firstname.lastname@example.org. Twitter: @TheloniousGoerz
Like what you’re reading? Support high-quality student journalism by donating here.