Models and Methods for Unsupervised Classification of Functional Data with Applications to Dynamic Survival Prediction via Landmarking

The aim of this project is to propose and develop new methods and models for the unsupervised classification of functional data. These methods will also leverage the functional nature of routinely collected longitudinal patient data to provide more accurate prognostic information for dynamic survival models. Techniques from functional clustering will be extended to account for the sparse and irregular nature of longitudinal data. The extracted information on heterogeneous biomarkers is then integrated into survival models that dynamically estimate survival curves within a landmarking framework. As new patient information becomes available, biomarker trajectories are clustered, and the resulting cluster assignments are used as prognostic predictors to improve survival prediction.