The proposed project aims to develop and validate robust predictive models for age-associated diseases by integrating multimodal biomarkers from routinely collected blood and whole-body MRI data within the UK Biobank, with external testing in a Swiss clinical cohort.
Primary research questions are:
(1) Can classical survival models and advanced deep-learning frameworks accurately predict individual risk of cardiovascular, metabolic, neurodegenerative, and oncological conditions in both UK and Swiss populations?
(2) How do Cox proportional hazards models compare to convolutional neural network-based architectures in predictive performance and clinical utility?
Specific objectives are to:
– Curate and preprocess common blood alongside imaging-derived phenotypes.
– Train and internally cross-validate risk models using UK Biobank participants.
– Externally validate model generalizability and calibration within a Swiss clinical cohort with equivalent multimodal data.
– Apply interpretability techniques (e.g., SHAP, feature-importance mapping) to elucidate key biological and imaging predictors of disease onset in a clinical context.
Integrating high-dimensional biomarker panels with imaging features offers unprecedented insight into subclinical disease phenotypes and inter-individual heterogeneity, informing personalized risk stratification. A systematic comparison of statistical versus deep-learning approaches will guide best practices for clinical deployment, while external validation across distinct populations ensures broad applicability and translatability.