Artificial Intelligence-Based Predictions in Neovascular AMD

Development and Application of Artificial Intelligence Models for Neovascular Age-related Macular Degeneration

Artificial intelligence, and particularly the subfield of deep learning, has the potential to identify features prognostic for individual patient outcomes. However, most advances in applying artificial intelligence to nAMD have focused on development and application of models to facilitate image analysis, particularly for automated segmentation, extraction, and quantification of imaging-based features from optical coherence tomography (OCT). Key artificial intelligence-based models for OCT image analysis and recent applications of artificial intelligence to nAMD are discussed herein.

Training, tuning, and testing of artificial intelligence-based algorithms typically require large, high-quality datasets. Nonetheless, in nAMD, multiple research groups have developed artificial intelligence-based algorithms using relatively few datasets, of which HARBOR and the Moorfields Eye Hospital real-world age-related macular degeneration (AMD) database stand out in the literature.

The phase 3 HARBOR trial (NCT00891735) assessed ranibizumab for 1097 patients with treatment-naïve nAMD, comparing two dosages and monthly and PRN treatment regimens.^[8,22] Notably, HARBOR was the first major clinical trial for nAMD to use spectral-domain OCT, which allows for high-sensitivity feature extraction.

Moorfields Eye Hospital, a tertiary referral retinal center in the United Kingdom, maintains a real-world database of electronic medical records and associated OCT images from patients with AMD treated with at least one ranibizumab or aflibercept injection from 2008 to 2018 and with at least 1 year of follow-up.^[23] Altogether, the Moorfields AMD dataset includes 8174 eyes of 6664 patients; a de-identified version of the segmentation results is openly available to the research community.^[23,24]

A key model for OCT image segmentation and disease classification, developed by De Fauw and colleagues,^[25] uses a deep learning-based framework with two independent networks to perform automated diagnosis of retinal diseases on OCT scans. This methodology has been applied to investigating imaging biomarkers and visual outcomes.^[24,26] Following this, another group developed a novel automated segmentation model using a convolutional neural network.^[27] This model was built using a large, real-world electronic medical records-based dataset from the United Kingdom, annotated by clinical experts with 13 of the most common AMD biomarkers on OCT, including IRF, SRF, and pigment epithelial detachment (PED).^[27]

The Notal OCT Analyzer^[28] and Medical University of Vienna artificial intelligence-based Fluid Monitor^[29] are fully automated tools for fluid detection and quantification on OCT images. These have facilitated quantitative measurements across multiple large datasets^[30] and been applied to questions investigating retinal fluid measurements and visual outcomes,^[17,30–35] particularly to more precisely quantify and map changes in IRF and SRF over time.

As an illustration, application of the Notal OCT Analyzer to a real-world dataset demonstrated that, by quartile, larger fluctuations in IRF, SRF, PED, central subfield thickness, and total fluid during the anti-VEGF maintenance phase were associated with worse visual acuity at 2 years.^[34] Other exploratory studies applying the artificial intelligence-based Fluid Monitor supported differential impact of IRF and SRF on vision outcomes. In both the HARBOR and FLUID trials, increased IRF, but not SRF, volumes in the central 1 mm were negatively associated with visual acuity [letters per 100 nl fluid, IRF: –4.00 and –2.84; SRF: +1.10 and +1.43 (not significant), respectively].^[32,33] Similarly, a stronger association of IRF than SRF with visual acuity was found by applying the De Fauw et al.^[25] methodology to the Moorfields AMD database.^[24]

Other studies have developed and applied artificial intelligence-based models to OCT image analysis for a diverse set of research questions, including determining whether visual acuity can be predicted from OCT;^[36] extracting higher-order features, such as ellipsoid zone integrity and subretinal hyperreflective material volume;^[37] facilitating correlational analysis among multiple features on OCT;^[35] comparing 'typical' nAMD with polypoidal choroidal vasculopathy;^[38,39] and clustering patients based on CNV features using unsupervised machine learning.^[39]