Book · 2026
Bayesian non-parametric fusion model for high-dimensional heterogeneous data and its application in precision medicine genomics
Proceedings of the 2nd International Conference on Artificial Intelligence, Digital Media Technology and Social Computing
Abstract
Precision medicine relies on the integration of multi-omics data to analyze disease heterogeneity, but the high-dimensional, heterogeneous and small-sample characteristics pose challenges for existing methods. This study proposes a Bayesian non-parametric fusion model based on hierarchical Dirichlet process and Indian buffet process, which adaptively integrates multimodal data (such as gene expression, methylation, copy number variation) through shared latent variables, and supports heterogeneous data types such as continuous, count and binary. Variational inference is used for efficient posterior approximation, accurately recovering the data structure and key features in simulated data; in TCGA multi-cancer data, the model automatically identifies cancer subtypes with significant survival differences (log-rank test p < 0.001), and the survival prediction performance (C-index) is superior to existing benchmark methods. This framework provides an interpretable, adaptive and uncertainty-quantifiable computational tool for multi-omics data fusion, which is conducive to promoting the clinical translation of precision medicine.