Equitable Artificial Intelligence for Glaucoma Screening with Fair Identity Normalization

Objective To develop an equitable artificial intelligence model for glaucoma screening. Design Cross-sectional study. Participants 7,418 optical coherence tomography (OCT) paired with reliable visual field (VF) measurements of 7,418 patients from the Massachusetts Eye and Ear Glaucoma Service between 2021 and 2023. Methods We developed fair identify normalization (FIN) module to equalize the feature importance across different identity groups to improve model performance equity. EfficientNet served as the backbone model to demonstrate the effect of FIN on model equity. The OCT-derived retinal nerve fiber layer thickness (RNFLT) maps and corresponding three-dimensional (3D) OCT B-scans were used as model inputs, and a reliable VF tested within 30 days of an OCT scan was used to categorize patients into glaucoma (VF mean deviation < -3 dB, abnormal glaucoma hemifield test (GHT) and pattern standard deviation (PSD) < 5%) or non-glaucoma (VF mean deviation ≥ -1 dB and normal GHT and PSD results). The area under the receiver operating characteristic curve (AUC) was used to measure the model performance. To account for the tradeoff between overall AUC and group disparity, we proposed a new metric called equity-scaled AUC (ES-AUC) to compare model performance equity. We used 70% and 30% of the data for training and testing, respectively. Main Outcome Measures The glaucoma screening AUC in different identity groups and corresponding ES-AUC. Results Using RNFLT maps with FIN for racial groups, the overall AUC and ES-AUC increased from 0.82 to 0.85 and 0.76 to 0.81, respectively, with the AUC for Blacks increasing from 0.77 to 0.81. With FIN for ethnic groups, the overall AUC and ES-AUC increased from 0.82 to 0.84 and 0.77 to 0.80, respectively, with the AUC for Hispanics increasing from 0.75 to 0.79. With FIN for gender groups, the overall AUC and ES-AUC increased from 0.82 to 0.84 and 0.80 to 0.82, respectively, with an AUC improvement of 0.02 for both females and males. Similar improvements in equity were seen using 3D OCT B scans. All differences regarding overall-and ES-AUCs were statistically significant (p < 0.05). Conclusions Our deep learning enhances screening accuracy for underrepresented groups and promotes identity equity. ### Competing Interest Statement The authors have declared no competing interest. ### Funding Statement This work was partially supported by NIH R00 EY028631, NIH R21 EY035298, Research to Prevent Blindness International Research Collaborators Award, Alcon Young Investigator Grant, and NIH P30 EY003790 ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: Ethics committee/IRB of Massachusetts Eye and Ear waived ethical approval for this work I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes In accordance with the policies of Massachusetts Eye and Ear, the data from this study cannot be publicly disclosed