Fig. 1

Overview of this study. (a) Dataset shift. This study used health checkup data from Hirosaki City in Aomori Prefecture, Japan, and Wakayama Prefecture, Japan, with dataset shift. The disease onset prediction model constructed from Hirosaki data has a lower prediction performance in Wakayama data than that of Hirosaki data due to the dataset shift. (b) Proposed method—out-of-distribution reject option for prediction; ODROP. In the proposed method, an out-of-distribution (OOD) detection model constructed from Hirosaki health checkup data first calculates the OOD score of each Wakayama health checkup data. The OOD score represents suitability as OOD data. Thus, data with an OOD score above a threshold are classified as OOD data (right side of OOD score histogram). Finally, a disease onset prediction model constructed from Hirosaki data predicts the in-distribution (ID) data, which are appropriate for prediction.