Artificial Intelligence-Based Predictions in Neovascular Age-Related Macular Degeneration

Daniela Ferrara; Elizabeth M. Newton; Aaron Y. Lee


Curr Opin Ophthalmol. 2021;32(5):389-396. 

In This Article

Predicting Neovascular Age-related Macular Degeneration Treatment Response With Artificial Intelligence

Several groups have begun to develop algorithms for predicting treatment response and treatment frequency needs. Within the parameters of our literature search, three studies to date specifically examined treatment response using artificial intelligence, with each approaching and defining response differently.

Two studies focused on the anatomic response to anti-VEGF treatment on OCT. The first developed a novel method with convolutional neural networks, using data from a real-world cohort at Second Affiliated Hospital of Xi'an Jiaotong University. The study found that effectiveness of anti-VEGF treatment on CNV or cystoid macular edema could be predicted with area under the curve of 0.81 using baseline OCT images.[40] However, 'effective' appeared to be a binary treatment response that was not clearly defined. In the other study, a conditional generative adversarial network was used to develop a deep learning model capable of generating posttreatment OCT images.[41] This model, trained on a real-world retrospective dataset from Konkuk University Medical Center, was designed to generate OCT images representing 1 month after completion of three monthly anti-VEGF loading doses. A model including baseline OCT, fluorescein angiography, and indocyanine green angiography images, rather than OCT images alone, performed best in its prediction of each of IRF, SRF, PED, and subretinal hyperreflective material.[41]

To explore predictive ability of quantitative OCT parameters for posttreatment visual outcomes, Fu et al.[26] applied De Fauw et al.'s[25] deep learning method to the Moorfields AMD database. Together, baseline visual acuity and OCT parameters had a predictive accuracy for 3 months post injection and 12 months post baseline of R2 = 0.49 and 0.38, respectively, which improved to R2 = 0.79 and 0.63 by incorporating previous treatment response (incremental visual acuity and OCT changes).

Finally, one group developed an end-to-end deep learning model for predicting treatment requirements for patients receiving anti-VEGF on a PRN regimen per investigator discretion; the specific patient population was not identified.[42] OCT images were analyzed based on previous models[29] for fluid quantification to exclude patients for whom model and investigator decisions disagreed on more than three noninjection events over 2 years. Based on longitudinal images, the model categorized patients as having 'low,' 'intermediate,' and 'high' treatment requirements (up to five, five to 15, and ≥16 injections, respectively). Although the model did not perform well classifying patients in the intermediate group, area under the curve of 0.85 and 0.81 was achieved in binary classifications of low versus all or high versus all treatment requirements.[42] However, this study did not ultimately correlate these treatment requirements with vision outcomes.[42]