In the JournalsPerspective

Algorithm predicts diabetic retinopathy progression

A deep learning algorithm developed by Genentech and Roche researchers predicts a patient’s diabetic retinopathy progression using color fundus photographs acquired at a single visit, according to a study.

“Our deep learning pilot study shows the potential for a tool that is highly accurate in predicting diabetic retinopathy progression at an individual patient level in a single doctor visit. Such a tool could help inform the patient’s treatment strategy — earlier treatment depending on the risk for fast progression — in order to protect and preserve their vision,” study author Zdenka Haskova, MD, PhD, medical director in clinical ophthalmology at Genentech, told Healio.com/OSN.

The retrospective study analyzed stereoscopic seven-field central fundus photographs obtained from eyes in the RIDE and RISE phase 3 studies at baseline. The eyes were treatment-naive to anti-VEGF therapy. Researchers generated algorithms from these data sets to predict worsening in untreated eyes from baseline over a period of 2 years.

Researchers then used seven-field central fundus photographs acquired from patients with diabetic retinopathy to train deep learning models to predict a two-step or more worsening on the ETDRS diabetic retinopathy severity scale over 2 years.

Worsening was predicted with an area under the curve of 0.68 ±0.13, a sensitivity of 66% and a specificity of 77% at 6 months. At 12 months, AUC was 0.79 ± 0.05, sensitivity 91% and specificity 65%. At 24 months, worsening was predicted with an AUC of 0.77 ± 0.04, a sensitivity of 79% and a specificity of 72%.

“This type of predictive algorithm could help patients get the individualized care they need, as their ophthalmologists could potentially tailor a patient’s treatment strategy knowing whether they are at risk for faster diabetic retinopathy progression,” Haskova said. – by Robert Linnehan

Disclosure: Haskova reports she is an employee and shareholder of Genentech.

A deep learning algorithm developed by Genentech and Roche researchers predicts a patient’s diabetic retinopathy progression using color fundus photographs acquired at a single visit, according to a study.

“Our deep learning pilot study shows the potential for a tool that is highly accurate in predicting diabetic retinopathy progression at an individual patient level in a single doctor visit. Such a tool could help inform the patient’s treatment strategy — earlier treatment depending on the risk for fast progression — in order to protect and preserve their vision,” study author Zdenka Haskova, MD, PhD, medical director in clinical ophthalmology at Genentech, told Healio.com/OSN.

The retrospective study analyzed stereoscopic seven-field central fundus photographs obtained from eyes in the RIDE and RISE phase 3 studies at baseline. The eyes were treatment-naive to anti-VEGF therapy. Researchers generated algorithms from these data sets to predict worsening in untreated eyes from baseline over a period of 2 years.

Researchers then used seven-field central fundus photographs acquired from patients with diabetic retinopathy to train deep learning models to predict a two-step or more worsening on the ETDRS diabetic retinopathy severity scale over 2 years.

Worsening was predicted with an area under the curve of 0.68 ±0.13, a sensitivity of 66% and a specificity of 77% at 6 months. At 12 months, AUC was 0.79 ± 0.05, sensitivity 91% and specificity 65%. At 24 months, worsening was predicted with an AUC of 0.77 ± 0.04, a sensitivity of 79% and a specificity of 72%.

“This type of predictive algorithm could help patients get the individualized care they need, as their ophthalmologists could potentially tailor a patient’s treatment strategy knowing whether they are at risk for faster diabetic retinopathy progression,” Haskova said. – by Robert Linnehan

Disclosure: Haskova reports she is an employee and shareholder of Genentech.

    Perspective

    Diabetic retinopathy is a leading cause of vision loss in working-age adults. One of the major issues with managing this disease is the ability to predict future progression. Recent data have highlighted that we can prevent and even regress retinopathy progression by applying anti-VEGF agents to these patients. And while many AI platforms to date have focused on detecting retinopathy, few have aimed at evaluating disease progression.

    The purpose of this study was to develop a scale for progression over a year period when provided just with a single color retinal fundus image. The authors developed a deep learning algorithm evaluating the seven standard fields and achieved an area under the curve of 0.79. Quite simply, we can use a single image to predict a two-step retinopathy progression at 6, 12 or 24 months. Thus, we can use this algorithm when commercially available to set follow-up periods more reliably for our patients.

    The one limitation is that the authors did use the RISE/RIDE prospective clinical trial data set, and thus validating this algorithm in other data sets might help to further solidify its power of prediction.

    • Rishi P. Singh, MD
    • Healio/OSN Board Member

    Disclosures: Singh reports he is a consultant to Genentech, Regeneron and Novartis.