This study explores the development of the General Demographic Pre-trained (GDP) model, a foundational model specialized for demographic attributes such as age and gender. We pretrained and evaluated the model using datasets with diverse diseases and population compositions. We explored how to transform tabular demographic inputs into effective latent embeddings by combining ordering and encoding methods to construct the GDP architecture. Consequently, GDP demonstrated generalizability across tasks, diseases, and populations. Specifically, the sequential ordering approach improved model performance in diseases where age and gender are important for risk stratification, and it also enhanced representational significance in datasets where demographic attributes have relatively low predictive value.
