Artificial intelligence may predict the course of geographic atrophy in individual patients, helping clinicians determine whether patients are benefiting from drug therapy.
Accurate predictions will become important if, as expected, drugs now in the pipeline are approved to slow the progress of the disease, said Tiarnán Keenan, MD, PhD, a staff clinician who researches age-related macular degeneration (AMD) at the National Institutes of Health in Bethesda, Maryland.
Researchers from Genentech presented these findings at the virtual European Society of Retina Specialists (EURETINA) 2021 meeting.
“The performance is the best I’ve seen from any group,” Keenan, who wasn’t involved with the study, told Medscape Medical News.
Some drugs in clinical trials now must be administered through intravitreal injections, which patients frequently find uncomfortable. But those drugs are not expected to improve vision, only to keep it from deteriorating. Keenan worries that patients might lose interest if there’s no way to tell if they are benefiting.
“If, after the treatment, the enlargement is slower than what was predicted, that would give some confidence,” he said. “The alternative would be to withhold treatment for 6 months to calculate the enlargement rate, then start the treatment, but then you’ve wasted 6 months.”
To create the programs, the Genentech researchers devised a training set with images of 1279 eyes created for a 2017 trial of its drug lampalizumab. That trial failed to show efficacy. But the researchers were left with a wealth of both fundus autofluorescence (FAF) and optical coherence tomography (OCT) images that showed how lesions had changed over time.
They fed these images to a convolutional neural network, a type of computer program, designed to search for patterns that correlate to different rates of growth. The program created an algorithm to predict growth, and the researchers tested it on a separate set of images of 443 eyes they had set aside from the same trial.
Using the combined FAF and OCT data, the program predicted the growth of the lesions with a score of 0.47 on a scale of 0-1 (R2 = 0.47), a correlation considered moderate. FAF by itself was equally accurate (R2 = 0.48), while OCT alone was less accurate (R2 = 0.36). By contrast, results with a simple linear model devised without the use of artificial intelligence achieved a weak correlation (R2 = 0.16).
“If you think about real-world applicability, if we can have an algorithm that uses only one modality instead of two, that’s less burdensome for patients and for healthcare,” said investigator Daniela Ferrara, MD, PhD, global development lead in ophthalmology at Genentech in South San Francisco, California. “So we’re very excited to see that fundus autofluorescence alone can give us very good prediction.”
The program can also make clinical trials more efficient by helping researchers select those patients whose lesions are growing fastest. That should make the results speedier as well.
Also, the researchers have programmed the software to generate heat maps that show which pixels are the most important for its predictions. That could provide new insights into the etiology of the disease. “AI can help us see beyond what the human eye can see,” Ferrara told Medscape Medical News.
Ferrara cautioned that the prediction algorithm still has to be validated in a real-world trial on a broader range of patients and images.
It might be enhanced as well by incorporating information about patients’ smoking history, diet, and genes, since these factors are associated with the rate of lesion growth, Keenan said. Several other teams are working on similar algorithms, some using such data, he said.
In a separate trial, Genentech researchers showed that a similar program using FAF and near infrared imaging was extremely accurate in an even more basic task: measuring the sizes of lesions. The difference between its measurements and those of two human graders was about the same as the difference between the measurements taken by the two human graders.
Measuring lesions is a laborious task, and the goal of this study was to find a way of speeding it up through automation, Ferrara said. “Algorithms are going to perform those tedious tasks better than humans because they don’t get tired and they don’t get distracted.”
Ferrara is an employee of Genentech. Keenan reported a patent pending for methods of predicting the course of age-related macular degeneration. The study was funded by Genentech.
European Society of Retina Specialists (EURETINA) 2021: Abstracts 8156 and 8165 Presented September 9, 2021.
Laird Harrison writes about science, health and culture. His work has appeared in national magazines, in newspapers, on public radio and on websites. He is at work on a novel about alternate realities in physics. Harrison teaches writing at the Writers Grotto. Visit him at lairdharrison.com or follow him on Twitter: @LairdH
Source: Read Full Article