Genetically Adjusted Optic Cup to Disc Ratio (CDR) Using a Two-Phase Training Deep Learning Model

Abstract

Purpose : The optic cup-to-disc ratio (CDR) is crucial in diagnosing glaucoma. While most studies to date use a CDR cut-off of 0.7 as indicative of glaucoma, clinically patients with glaucoma often have a CDR less than 0.7 and vice versa. The ability to differentiate pathologic from physiologic optic disc cupping is critical for improving glaucoma risk assessment and diagnostic accuracy. Here, we aim to adjust CDR for the patients based on their genetic characteristics.

Methods : We calculated a CDR polygenic risk score (PRS) for 6575 patients with ocular data in a Mass General Brigham (MGB) clinical biobank based on primary open-angle glaucoma (POAG) independent loci associated with CDR in a prior large genome-wide association study [1]. We trained a deep learning model with two-phase training: in the first phase the model is a classifier predicting patient status based on ICD codes (0, non-glaucoma eye or 1, glaucomatous eye) using the PRS and raw CDR values (Figure 1a1). In the second phase, the model's last layer is replaced with another layer to calculate the adjusted CDR. For the patients whose CDRs comply with the 0.7 criterion, their PRS and CDRs (i.e., raw CDRsL Figure 1a1) are input to the model and their CDRs are the outputs of the model (i.e., adjusted CDRs: Figure 1a2) in this phase.

Results : The cohort consisted of 53% females with an average age of 68.2 yrs (14.33 yrs/std). The majority were of European (86.55%) ancestry, while 6.18% were Hispanic, 5.46% African, and 1.81% Asian. There is a high-class imbalance in the dataset in which 85.61% of patients had no glaucoma ICD code, and 14.39% had at least one H40 or 365 ICD code. The model was tested with 399 samples (197 non-glaucoma, and 202 glaucomatous) which were manually reviewed by two ophthalmologists. Using the raw CDR values, 7.42% of non-glaucomatous eyes were misclassified as glaucomatous while 26.54% of glaucomatous eyes were misclassified as normal (Figure 1b). Adjusted CDR however achieved for comparison, the f1 score for the non-glaucoma eye class was 0.95 and 0.94 for the glaucoma eyes with a macro average is 0.94 (Figure 2a). The R2 score in the regression mode for the test set was 0.9989 (Figure 2b).

Conclusions : Using a CDR PRS and two-phase training deep learning, we were able to adjust CDR values for the patients to represent more accurate clinically verified disease.

This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.