Peripheral Fields Key in Predicting Diabetic Retinopathy

Donavyn Coffey

November 22, 2019

A new machine learning model can predict how quickly diabetic retinopathy will progress in individual patients. Unexpectedly, the algorithm relies on information clinicians rarely consider in routine diabetic retinopathy screening, according to a study published online in NPJ Digital Medicine. 

Using images from a single doctor's visit, the algorithm, developed by researchers at Roche and Genentech, can determine if a patient with diabetes is likely to experience vision loss in the next 1 to 2 years. Unlike existing screening methods that rely on images of the fovea and optic nerve, the new system found equally important predictors in the peripheral retinal fields.

Researchers trained the model using diabetic retinopathy severity scores provided by masked, human reading center graders at 6, 12, and 24 months after the baseline visit. Images for the study were from the RISE and RIDE clinical trials, taken using seven-standard field imaging. Researchers tested the model on baseline images of eyes with untreated diabetic retinopathy to see if it predicted a two-step or more worsening on the Early Treatment Diabetic Retinopathy Severity Scale within 2 years.

"We found that the main predictive contribution came from the peripheral retinal fields that encompass areas of the retina far from both the fovea and optic nerve," writes Filippo Arcadu, PhD, from Roche Informatics in Basel, Switzerland, and colleagues. Funding for the study was provided by Roche and Genentech.

Specifically, when the model was trained using images from all seven fields, the area under the curve (AUC) was 0.68 at 6 months and 0.77 at 24 months. When the researchers restricted training to images of the fovea and optic nerve only, the AUC fell to 0.62 and 0.69, respectively (P = 0.0486 and 0.0023). 

Though typical exams may not look at the periphery, this research confirms the importance of looking at the entire retina, Zdenka Haskova, MD, PhD, a senior study author and clinical ophthalmologist at Genentech, told Medscape Medical News. If the algorithm is coupled with more recent imaging technologies like ultra-widefield photography it may be possible to identify fast-progressing diabetic retinopathy even earlier.

"One thing that needs to be studied is whether lesions that are on the macula and disc predict periphery lesions," says Paolo Silva, MD, a professor of ophthalmology at Harvard's Joslin Diabetes Center, Boston, Massachusetts, who was not involved in the study. "If [they don't], then the periphery is critical in detecting diabetic retinopathy."

Although the study's data set used seven-field imaging, which is not practical in a population-based diabetic retinopathy screening setting, Haskova says that part of the validation and expansion of the model will be to test the algorithm in a larger, real-world data set and make sure it can operate on ultra-widefield images. 

Silva, who has published extensively on this novel imaging technology, said he thinks the ultra-widefield images should work fine in the algorithm because they provide even more data than seven-fields imaging.

"With limited healthcare resources, it is important to know which patients need the most attention," Haskova said. Current screening methods are insufficient, given the pandemic size of the global diabetic population (approximately 425 million) and the fact that diabetic retinopathy is largely asymptomatic until vision loss begins to set in.

If an algorithm can identify higher-risk patients using images taken at a single doctor's visit, "then you could direct limited healthcare, and, for example, schedule their visits more frequently," Haskova said.

The artificial intelligence system, once validated, could offer physicians a way to better use all the information they have. "If you integrate all of that together — EMR, systemic risk factors, retinal risk factors — you can potentially provide each patient with real-time decision support," Silva said.

NPJ Digit Med. 2019;2:92. Full text

Follow Donavyn Coffey on Twitter: @donavyncoffey

Follow Medscape on Facebook, Twitter, Instagram, and YouTube.

Comments

3090D553-9492-4563-8681-AD288FA52ACE
Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.
Post as:

processing....