Fig. 8: Comparison of Ci-SSGAN and ICD code–based labeling across glaucoma subtypes and non-glaucoma cases in terms of Accuracy, F1 Score, AUC-ROC, and AUC-PR. | npj Digital Medicine

Fig. 8: Comparison of Ci-SSGAN and ICD code–based labeling across glaucoma subtypes and non-glaucoma cases in terms of Accuracy, F1 Score, AUC-ROC, and AUC-PR.

From: Clinically informed semi-supervised learning improves disease annotation and equity from electronic health records: a glaucoma case study

Fig. 8

Each bar represents performance for a specific class, with the “Overall” category summarizing all classes. Ci-SSGAN uses both labeled and unlabeled data with clinical context, whereas ICD code labels rely solely on diagnosis codes from medical records. Improvements were most pronounced in challenging subtypes such as primary angle-closure glaucoma (ACG/S), open-angle glaucoma (OAG/S), and Non-GL cases. AUROC and AUCPR calculated using one-vs-rest strategy with macro-averaging across six glaucoma classes.

Back to article page