The spectrum of major chronic diseases-including metabolic diseases (e.g., T2D, NAFLD), respiratory diseases (e.g., COPD), cardiovascular diseases (e.g., ASCVD), and major malignancies-is the leading cause of global mortality, characterized by substantial heterogeneity. Traditional epidemiology is constrained by single-modal perspectives and confounding, failing to capture the continuum from molecular variation to clinical outcomes. Addressing this, we leverage the UK Biobank to construct a multi-dimensional atlas of major chronic diseases integrating multi-omics and radiomics.
We follow a three-tiered framework: “Multi-dimensional Characterization,” “Causal Inference,” and “Deep Learning Prediction.” First, we transcend binary classifications by integrating GWAS, proteomics, metabolomics, and multi-modal MRI/CT data (brain, heart, abdomen, chest). Using unsupervised algorithms (e.g., GMM), we identify cross-disease latent subtypes defined by specific genetic and molecular profiles, laying a foundation for stratified analysis.
Second, we employ Two-Sample Mendelian Randomization (MR) to validate causal relationships between lifestyle and these disease phenotypes, minimizing confounding bias. Concurrently, we apply mediation analysis to quantify the effects of multi-omics biomarkers along the “exposure-phenotype-outcome” pathway, revealing biological targets for intervention.
Finally, utilizing deep learning, we construct a cross-disease multi-modal fusion model. By leveraging CNNs/ViTs to extract radiomic features combined with Polygenic Risk Scores (PRS), we build an end-to-end engine to predict all-cause mortality and major adverse events. We evaluate the incremental value over traditional scores and verify robustness across subgroups. In summary, this study drives a shift from “association” to “causality” and “precision prediction” in chronic disease research, supporting personalized health strategies
