Type 2 diabetes is a common lifestyle-related disease that is likely to be prevented by detecting early signs. Health checkups are the promising way to reduce the risk of lifestyle-related diseases since they cover the yet-to-be-sick population. From large-scale diachronic tabular health checkup datasets, we aimed to predict future hemoglobin A1c (HbA1c), a standard indicator of diabetes. Based on the biological consideration that factors are specific to each disease and factors common across multiple diseases as well, we developed a Transformer-based branched neural network model containing a module emphasizing the differences between tasks. We applied multi-task learning with the prediction of HbA1c and creatinine to this model.