Fig. 2
From: Opportunistic screening of type 2 diabetes with deep metric learning using electronic health records
T2D onset prediction performance. (a) Temporal performance of the DML, LR, Glycemic, and Wilson models on AoU PopControl data with censor periods ranging from 10 years to 0 years before diagnosis (95% CIs over 500 bootstrap iterations). (b) Bar plot of 2-year T2D onset prediction. AUROC of the DML with LR-baselines (LR, Risk-Factors, Glycemic), deep learning baselines (SCARF, TabTransformer, CVAE, ConvAE), and dimensionality reduction baselines (PCA, UMAP). 95% CIs computed over 500 bootstrap iterations. (c) Transfer performance of MGB-trained DML and LR models evaluated on AoU data. (d) Latent Space Representations from the AoU DML model are visualized through dimensionality reduction with UMAP.
