This project aims to develop robust and interpretable machine learning models for predicting cardiometabolic and mental health outcomes using multimodal data from the UK Biobank. Cardiovascular disease, diabetes, and common mental disorders remain major contributors to morbidity and mortality, yet existing risk models often rely on limited data modalities and perform poorly across diverse population subgroups.
The research will address three key questions:
(1) How can multimodal data-including demographics, lifestyle factors, clinical measurements, imaging, and genetic information-be effectively integrated to improve disease risk prediction?
(2) How does missing or heterogeneous data affect predictive performance, and how can models be designed to remain robust under realistic data limitations?
(3) Can uncertainty-aware and interpretable models provide clinically meaningful insights into individual and population-level risk factors?
Methodologically, the project will develop statistical and machine learning approaches, including probabilistic models and deep learning methods, to handle missing modalities, domain shift, and population heterogeneity. Model performance will be evaluated across demographic subgroups to assess robustness and potential biases. Emphasis will be placed on transparency, uncertainty quantification, and reproducibility.
The scientific rationale is to improve the reliability and equity of data-driven risk prediction in large population cohorts. By leveraging the scale and richness of UK Biobank data, the project seeks to generate evidence that can inform early risk stratification, support preventive strategies, and contribute to improved public health decision-making. All analyses will be conducted for health-related research in the public interest, with no attempt to identify individual participants.