Discriminative latent representation harmonization of multicenter medical data

Data harmonization is critical for establishing generalizable model on multicenter medical data. Traditional data harmonization strategies aim to align data distributions from different sources, but often lack mechanisms to learn hidden complementary and discriminative manifestations from multicenter data. To this end, we proposed a methodology for harmonizing multicenter data by matching their first and second order statistics in a shared space, which is framed in an optimization architecture to learn harmonized and discriminative latent features for downstream classification modeling. The developed method integrated representation learning, feature dimension reduction and selection within a unified framework. Several relational regularizations such as data attribute preservation and feature-task correlation have been explored and incorporated to encourage learning potential associations inherent in multicenter data. Extensive evaluations on three independent clinical datasets have demonstrated the efficacy of the proposed method in producing harmonized and distinguishing data for multicenter medical prediction modeling.