A1192
Title: Unlocking precision and smart health: Insights from genetics, medical imaging, and multimodal AI integration
Authors: Yi-Jia Huang - Academia Sinica (Taiwan)
Yun-Jer Shieh - Academia Sinica (Taiwan)
Chun-houh Chen - Academia Sinica (Taiwan)
Hsin-Chou Yang - Academia Sinica (Taiwan) [presenting]
Abstract: Genetic variation and medical imaging capture complementary dimensions of human health, yet their integration for mechanistic insight and risk prediction remains limited. A scalable multimodal framework that unifies genetic, imaging, and clinical data for integrated discovery and prediction is presented. Polygenic risk scores were constructed using internal and cross-ancestry genetic data, while imaging representations were derived via two complementary strategies: pixel-level modeling of raw DICOM images and feature-based modeling of image-derived phenotypes (IDPs). These components were integrated using multimodal fusion models combining genetic, imaging, and demographic variables. The framework was applied to type 2 diabetes (T2D) as an illustrative example in up to 131,048 Han Chinese participants from the Taiwan Biobank, integrating genome-wide imputed genotypes (~9.8 million variants), large-scale imaging data, 152 IDPs, and clinical variables. GWAS and QTL analyses, together with cross-phenotype and network-based approaches, characterized shared genetic architecture, while bidirectional Mendelian randomization inferred causal relationships. Multimodal integration consistently outperformed single-modality approaches. Pixel-based modeling improved scalability by reducing reliance on manual annotation, and integrative modeling enhanced both predictive performance and biological interpretability. A generalizable paradigm for precision health is established.
