Fig. 6
From: Out-of-distribution reject option method for dataset shift problem in early disease onset prediction
Dataset shift in diabetes onset within one-year records for diabetes onset prediction model. (a) SHAP clustering for diabetes onset within one-year records in Wakayama health checkups. This figure shows a hierarchical clustering analysis using SHAP values from a 1-year diabetes onset prediction model for individuals from Wakayama health checkup data who developed diabetes within 1 year. A colormap represents the magnitude of the SHAP values calculated by the prediction model, with the vertical axis listing the Wakayama health checkup data of individuals who developed diabetes within one year. The horizontal axis without an index column shows the names of each examination item used in the prediction model, whereas an index column is IDs and OOD labels based on the VAE reconstruction loss threshold at the rejection rate of 31.1%, where AUROC was maximized in the rejection curve. (b) HbA1c Levels in one-year diabetes onset Wakayama ID and OOD data (mean ± std). The HbA1c value, which showed the most pronounced pattern differences between ID and OOD in SHAP Clustering, was presented as mean ± std for both ID and OOD data.
