Equitable artificial intelligence for glaucoma screening with fair identity normalization.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma is a primary cause of irreversible blindness worldwide.
• Racial and ethnic minorities are disproportionately affected by glaucoma.
• Current deep learning models may exhibit performance disparities across diverse populations.

Purpose of the Study:

• To develop a novel module for enhancing fairness in deep learning models for glaucoma detection.
• To improve the equity of model performance across different racial and ethnic groups.
• To address the challenge of unequal performance in AI-driven diagnostic tools.

Main Methods:

• Utilized optical coherence tomography (OCT) measurements for glaucoma and non-glaucoma classification.
• Developed and implemented a fair identify normalization (FIN) module.
• Employed equity-scaled area under the receiver operating characteristic curve (ES-AUC) to measure performance equity.

Main Results:

• The FIN module improved overall AUC from 0.82 to 0.85 for racial groups, with AUC for Black individuals increasing from 0.77 to 0.82.
• ES-AUC increased from 0.77 to 0.81 for racial groups after applying FIN.
• For ethnic groups, overall AUC rose from 0.82 to 0.84, and ES-AUC from 0.77 to 0.80, with Hispanic individuals' AUC improving from 0.75 to 0.79.

Conclusions:

• The fair identify normalization (FIN) module significantly enhances the equity of deep learning models for glaucoma detection.
• FIN improves diagnostic performance across diverse racial and ethnic groups, mitigating existing disparities.
• This approach holds promise for more equitable AI-based healthcare solutions in ophthalmology.