Fig. 4: Prediction of T1D using machine learning of variant genotypes.
From: Genetic association and machine learning improve the prediction of type 1 diabetes risk
Receiver operating characteristic curves assessing the accuracy of predicting T1D from unaffected individuals using T1GRS and GRS2. The AUC for T1GRS is colored purple, while the existing GRS2 is colored red. a–c, The AUCs for T1GRS-cov and GRS2 are shown in the discovery dataset for all variants (a), MHC-only variants (b) and non-MHC genome-wide variants (c). d–f, Validation using T1GRS-var was performed in the NIH AoU Research Cohort for all variants (d), MHC-only variants (e) and non-MHC genome-wide variants (f). g–i, Validation in the nPOD cohort comparing T1GRS-var to GRS2 for all variants (g), MHC-only variants (h) and non-MHC genome-wide variants (i). P values comparing the predictive ability of GRSs for a–i were calculated using a two-sided DeLong test. j, The difference in false-negative rates (GRS2 − T1GRS-cov) for the discovery cohorts. P values were calculated by a two-sided McNemar’s test for each category. k, The difference in false-positive rates (GRS2 − T1GRS-cov) for the discovery dataset. P values were calculated by a two-sided McNemar’s test for each category.
