Roche-Genentech sees future for AI in retinal disease prevention

VANCOUVER, British Columbia — Roche-Genentech is working toward a future in which physician and machine come together to improve outcomes for patients.

Executives and researchers shared their vision here at a company-sponsored event during the Association for Research in Vision and Ophthalmology meeting for the potential of using technology for not only personalized treatment but disease prevention.

“It’s very different than what we’ve done, which is enroll patients with fairly advanced disease in clinical trials,” Jill Hopkins, MD, FRSC, global head of personalized health care – ophthalmology, Roche-Genentech, said at the event. “We’re trying to dial that way back and say: Can we detect earlier disease? Can we prevent vision loss from actually occurring?”

The company invested 9,000 person-hours over the course of 2 years to take its compendium of clinical trial data in the areas of ophthalmology, neuroscience and oncology and convert it to an annotated, curated format to which artificial intelligence techniques could be applied, Hopkins said.

The objective in eye care is to provide patients who have lost vision in one eye a predicted risk that they will lose vision in the fellow eye, along with a prevention strategy, she said.

“We’d like to image someone and say they have a very good likelihood of developing a neovascular membrane or going blind in that eye next year, and here’s what a prevention strategy might look like,” Hopkins said. “You could take patients with diabetic retinopathy who are at very high risk of progressing and start to treat them earlier and prevent vision loss from ever occurring.”

Anti-VEGF treatment “does a good job,” she said, “but we’re not good at predicting who needs what. Someone might need three injections and do just great with their age-related macular degeneration; someone else might need a monthly injection for the next 2 years to have the best visual outcomes.”

An agent in the company’s pipeline, faricimab, “has a dual mechanism of action, with not just anti-VEGF,” Hopkins said. “We might be able to identify patients who need combination therapy.

“Our port delivery system in phase 3 clinical trials would need a 6-month refill as opposed to a monthly or bimonthly injection,” she continued. “Could you predict patients who have that high anti-VEGF need and know they’re a good candidate for the port delivery system at the time of their diagnosis as opposed to waiting 6 months, having six injections and waiting to see how you do?”

Hopkins noted that AMD clinical trial outcomes are not being seen in the real world.

“The burden of regular therapy just can’t be sustained by patients or providers,” she said.

The company sees a future in which consumers can have a retinal scan taken at their local drug or grocery store and learn whether or not they have eye disease.

“Once you are diagnosed, you could look at an individualized treatment plan so you would know what to expect over the next year — the number of visits, treatments,” Hopkins said. “We’d love to see a world where patients could have tools at home or apps on their phone so they wouldn’t have to even be seen in a medical setting.”

A study published in March by Arcadu and colleagues showed that deep learning is able to detect disparity of diabetic macular edema from color fundus photos, Jeffrey Willis, MD, PhD, assistant medical director, ophthalmology, Genentech, said at the event.

“The gold standard [diagnostic for] DME is OCT,” Willis said. “This technology is not frequently used in telescreening, but color fundus photography, 2-D images are, and it’s very hard to detect the level of swelling.”

Willis said he and his colleagues used the internal data set from the RIDE and RISE trials.

“For every color fundus photo we had an OCT, so we had powerful data,” he said. “We used 18,000 fundus photos, put them through the deep learning algorithm and looked at three questions.”

The questions were whether AI could detect a color fundus photo with the thickness of 250 µm, the thickness of 400 µm and the actual thickness of the macula from the color fundus photos.

“The best models were able to detect, at a 97% accuracy, the color fundus photos that had 250 µm thickness value and a 94% accuracy of detecting color fundus photos with a 400 µm level thickness,” Willis said. “This is meaningful for future teleophthalmology screening programs. For the third question, detecting the actual thickness, the correlation was very high.”

Willis said that the deep learning algorithm picked up on elements that were biologically plausible, “focusing on the macula, changes in the retinal vessel caliber and changes in the hemorrhages. This was reassuring to us.”

He added: “We don’t know if this will be validated in the real world.”

Moving onto predictive outcomes, Filippo Arcadu, PhD, senior scientist, image analysis specialist, Roche, said researchers looked at predicting how patients would respond to therapy after 6, 12 or 24 months using data from 380 eyes from the RISE and RIDE trials.

“We saw that it was feasible that, using just baseline therapy, it was possible to predict if a patient would respond well to therapy,” Arcadu said.

A limitation to the study was its size, at 380 patients, he said.

“We need to ... come to numbers in the thousands and confirm that it’s still possible,” Arcadu said. – by Nancy Hemphill, ELS, FAAO

Reference:

Arcadu F, et al. Invest Ophthalmol Vis Sci. 2019;doi:10.1167/iovs.18-25634.

Disclosures: Arcadu reports he is employed by Roche. Hopkins reports she is employed by Roche-Genentech. Willis reports he is employed by Genentech.