Neural Networks for Time-To-Event Analysis

Type: Master Thesis

Student: Pierre Springer

Supervisor: Sebastian Pölsterl, Christian Wachinger

Status: Finished, 29.11.2019

 

Abstract:

Survival analysis is a significant research area for many real-world problems in medicine and science. While Cox-Proportional Hazard models were the standard choice for many years, several new algorithms evolved recently. In this thesis, we present novel survival networks which do not only allow to make time-dependent predictions but allow personalised recommendations for patients with Alzheimer’s disease.

For our experiments, we work with data provided by the Alzheimer’s Disease Neuroimaging Initiative. We review conventional and state-of-the-art survival analysis methods. For the representation of non-linear structures, we implement survival net- works, combining survival analysis and neural networks. We analyse these survival networks using Shapley Values, which is a theory coming from cooperative game theory and estimate the contribution of a feature to the total outcome.

This work concludes with a discussion of the results obtained from all experiments. For this, we examine not only Adni data but other medical real-world datasets. We generate synthetic data to specify the behaviour of our method under precisely defined constraints. This shows the strengths, but also the limitations of our algorithms. Finally, we show how our results can be used to find the optimal time point for each patient for screening based on one baseline visit.