Equitable Deep Learning for Glaucoma Progression Forecasting

Abstract

Purpose : Blacks have more than doubled glaucoma prevalence than other races, while deep learning performance for Blacks is typically worse due to data scarcity for minorities. We aim to develop an equitable deep learning model for glaucoma progression forecasting.

Methods : We selected reliable optical coherence tomography (OCT) scans from patients with at least 5 reliable 24-2 VFs over at least 4 years. We calculated 4 VF progression outcomes: (1) mean deviation (MD) progression: MD slope < 0 and p-value < 0.05; (2) VF index (VFI) progression: VFI slope < 0 and p-value < 0.05; (3) total deviation (TD) pointwise progression: at least 3 TD locations with TD slope ≤ -1 dB/year and p-value < 0.05; (4) MD fast progression: MD slope ≤ -1 dB/year and p-value < 0.05. We developed an equitable deep learning model guided by a fair loss to predict VF progression using baseline retinal nerve fiber layer thickness maps (Figure 1). The fair loss reweighted training loss based on effective sample sizes for each race. Our models were trained and tested using 2/3 and 1/3 of the dataset with patient-level separation, respectively. T-test with bootstrapping sampling was used to compare the performance of our models with and without fair loss guidance measured by the area under the receiver operating characteristic curve (AUC).

Results : 14,928 OCT scans from 4,726 eyes of 2,850 patients (age: 62.1 ± 12.8 years; MD: -3.0 ± 4.1 dB; White: 76.9%; Black: 14.6%; Asian: 8.5%) were included. The progression prevalence for MD progression, VFI progression, TD pointwise progression, and MD fast progression were 9.8%, 10.9%, 11.3%, and 2.5% with a median follow-time of 6.6 years, respectively. For MD, VFI and TD pointwise progressions, AUCs for Asians (Figure 2) increased from 0.68 to 0.75, 0.76 to 0.78 and 0.80 to 0.82 (p < 0.001), respectively. For all four progression outcomes, AUCs for Blacks increased from 0.82 to 0.84, 0.76 to 0.86, 0.84 to 0.89 and 0.83 to 0.89 (p < 0.001). AUC for Whites improved for MD progression prediction from 0.80 to 0.83 and stayed uncompromised for the other three progression outcomes. Our model with fair loss learned features with greater distances between racial groups than the model without fair loss.

Conclusions : Our equitable deep learning model improved performance for Asians and Blacks and did not compromise performance for Whites, which can potentially reduce health disparity in medical artificial intelligence.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Figure 1

 

Figure 2