OR WAIT null SECS
Artificial intelligence (AI) and deep learning (DL) can predict which patients with diabetic retinopathy (DR) will progress the fastest based on a single color fundus photograph, researchers say.
Arcadu et al. said their results “highlight the importance of the predictive signal located in the peripheral retina fields, not routinely collected for DR assessments, and the importance of microvascular abnormalities.”1
In-person expert examinations are impractical and unsustainable given the pandemic size of the diabetic population. As such, AI may offer a solution to this conundrum. DL, and specifically, deep convolutional neural networks (DCNNs), can be used for an end-to-end assessment of raw medical images to produce a target outcome prediction, the authors wrote.
Study design and results
The authors noted they wanted to assess the feasibility of developing DCNNs operating on seven-field color fundus photographs that can predict the future threat of significant DR worsening at a patient level over a span of 2 years after the baseline visit.
The study was based on a post-hoc, retrospective analysis of eyes with DR from the phase III RIDE and RISE trials (NCT00473382 and NCT00473330, respectively). The group used seven-field color fundus photographs from a single clinic visit to predict DR progression according to the Early Treatment Diabetic Retinopathy Study Diabetic Retinopathy Severity Scale (ETDRS DRSS) at 6, 12, and 24 months. With a high rate of loss to follow-up in this population, a single visit may be the only opportunity clinicians have to evaluate patients.2
At 6, 12, and 24 months, the DL model was able to accurately predict a 2-step or more ETDRS DRSS worsening in patients with DR, with an area under the curve (AUC) of 0.68 ± 0.13, 0.79 ± 0.05, and 0.77 ± 0.04, respectively. A total of 529, 528, and 499 patients had their 7 color fundus photograph fields captured on months 6, 12, and 24, respectively, and the majority of CFP images were high quality. Results at month 6 were weaker compared with months 12 and 24 because few patients progressed within the 6-month timeframe. A 2 years, the rates of 2-step or more worsening in sham study eyes and fellow eyes from baseline were 9.6% and 11.7%, respectively.
In addition to testing the feasibility of using AI to create an accurate algorithm that forecasted progression, the researchers also wanted to evaluate the predictive signal of DR progression in F1 and F2, generally considered the most important fields in a standard ophthalmoscopy exam, compared with the peripheral fields (F3, F4, F5, F6, and F7). Interestingly, researchers found that the peripheral retinal fields had more predictive significance than the central fields. As such, the researchers recommend that physicians assess both the central and peripheral retina rather than the central retina alone.
This algorithm narrows disease progression to a single individual, whereas currently it is only possible to assess DR progression risk in groups of patients. According to the authors, a second crucial finding of this study was that any imaging-based diagnostic/predictive AI tool for DR should contemplate the inspection of both central and peripheral retina, instead of being limited to the use of color fundus photographs centered around the fovea or optic nerve.
The authors did note, however, that a limitation of their study is that the DCNNs were developed and evaluated only data from the RISE and RIDE studies. “This means that, at this time, our work is only applicable to clinical trial populations within a similar range of pre-specified eligibility criteria,” the authors wrote. “Validation with datasets acquired in the real world will be essential to ensure that these results are reproducible and applicable to the broader DR population.”
The group also postulated that identifying fast-progressing patients could help in the development of clinical trials for new treatments, concluding that “by enriching the clinical trial population with fast DR-progressing individuals, such an AI-based recruitment strategy would increase the chances of success for clinical trials of novel drugs designed to address the unmet need of those members of the early DR population at the greatest risk of progression and vision loss. This is particularly important considering the rising global prevalence of DR and its potential impact on healthcare systems and society.”1
Finally, the group predicted deploying tools such as this one may save the U.S. healthcare system upwards of $624 million annually.
1. Arcadu F, Benmansour F, Maunz A, et al. Deep learning algorithm predicts diabetic retinopathy progression in individual patients. NPJ Digit Med. 2019;2:92.
2. Obeid A, Gao X, Ali FS, et al. Loss to follow-up in patients with proliferative diabetic retinopathy after panretinal photocoagulation or intravitreal anti-VEGF injections. Ophthalmology. 2018;125:1386-1392.