Fairness-aware deep learning for predicting visual field loss from optical coherence tomography

Standard automated perimetry is the clinical standard for measuring visual field (VF) loss in glaucoma, but it is subjective, time-consuming, and variable across patient visits. In contrast, optical coherence tomography (OCT) provides fast and reproducible structural measurements, creating an opportunity to estimate functional visual field loss directly from imaging. We developed and evaluated a fairness-aware volumetric deep learning framework that integrates an adaptive fairness feedback (AFF) mechanism with three-dimensional (3D) OCT modeling to jointly improve predictive accuracy and demographic equity for 52-point visual field prediction.

Using 3300 paired OCT-VF samples from the Harvard-GF dataset with a patient-level split, we trained a 3D ResNet-18 model incorporating demographic embeddings and an AFF mechanism that dynamically adjusts learning rates based on training subgroup error disparities. Model selection was performed using an internal validation subset of 300 samples. Performance was evaluated using mean absolute error (MAE), root mean squared error (RMSE), $R^2$ , equity-scaled metrics, subgroup disparity analysis, pointwise spatial evaluation, and bootstrap resampling.

The fairness-aware model with demographic embeddings and AFF achieved RMSE 5.21 dB, MAE 3.32 dB, and $R^2$ 0.56 on the held-out test set. Compared with a 3D baseline without demographic embedding or AFF, the full model reduced worst-group MAE from 3.90 to 3.53 dB and MAE disparity from 0.54 to 0.34 dB across racial subgroups, corresponding to a 37% reduction in racial MAE disparity while simultaneously improving overall accuracy.

Fairness-aware volumetric modeling with AFF improved OCT-based visual field prediction accuracy and reduced racial performance disparities. Ethnicity-based results should be interpreted cautiously because the Hispanic subgroup was small.