TransFair: Transferring fairness from ocular disease classification to progression prediction

Area of Science:

• Ophthalmology
• Medical Artificial Intelligence
• Computer Science

Background:

• Artificial intelligence (AI) in disease classification offers cost savings and improved access to care.
• Concerns exist regarding AI fairness, particularly its disproportionate impact on underprivileged groups.
• Existing methods for AI fairness in classification may not extend to disease progression prediction due to limited diverse longitudinal data.

Purpose of the Study:

• To introduce TransFair, a novel method for enhancing demographic fairness in ocular disease progression prediction.
• To ensure fairness is preserved when transferring knowledge from a fair disease classification model to a progression prediction model.

Main Methods:

• Trained a fairness-aware EfficientNet (FairEN) model using extensive ocular disease classification data.
• Adapted the FairEN model to a fair progression prediction model using knowledge distillation.
• Minimized latent feature distances between classification and progression prediction models to preserve fairness.

Main Results:

• Evaluated TransFair using 2D and 3D retinal images for ocular disease classification and progression prediction.
• Demonstrated that TransFair effectively enhances group fairness in predicting ocular disease progression.
• FairEN showed effectiveness in fairness-enhanced ocular disease classification.

Conclusions:

• TransFair successfully enhances demographic fairness in ocular disease progression prediction.
• The knowledge distillation approach effectively transfers fairness from classification to progression prediction.
• TransFair offers a promising solution for equitable AI in medical prognostics.