Multivariate survival analysis for predicting age-dependent disease risk from plasma proteomes

Ageing is the leading risk factor for various different diseases such as cardiovascular diseases, neurodegenerative diseases or cancers. In times of an ageing world population, age-related diseases will contribute significantly to the global health burden. In order to understand the underlying mechanisms of ageing and age-related disease risk and to predict age-dependent disease risk, we need good models of ageing. In this thesis, we use variational inference to develop a multivariate mixed accelerated failure time model. It is the first multivariate survival analysis model that is applied to more than 100 outcomes simultaneously. This model is able to identify disease development mechanisms that are shared across different diseases and learns how already existing comorbidities influence the risk of getting other diseases. We apply this model on the UK Biobank plasma proteomics data set and show that it outperforms established models, such as logistic regression, a Cox proportional hazards model and a multivariate fixed model on disease risk prediction with respect to the c-index. We further identify six clusters of diseases that are driven by similar proteomic mechanisms. An analysis of the most important proteins in each of these clusters reveals many of these proteins are already known to be part of the ageing proteome, such as NEFL, GFAP, NT-proBNP, COL9A1, HAVCR1, PAEP and TSPAN1. A further analysis of the comorbidity structure reveals seven disease clusters. Lastly, we identify a small set of 10 proteins that are as predictive for disease risk as the full set of 3000 proteins, which is helpful in clinical settings where only few proteins can be measured. This model is an alternative to more established ageing models such as age clocks, which model the difference between biological and chronological age, because it predicts disease risk directly. Additionally, unlike disease agnostic models, it models processes driving disease development simultaneously. We demonstrate that this multivariate approach is more predictive for future disease risk than predicting each disease separately.