Title: Time-to-event prediction with neural networks and Cox regression
Authors: Haavard Kvamme - University of Oslo (Norway) [presenting]
Abstract: New methods for time-to-event prediction are proposed by extending the Cox proportional hazards model with neural networks. Building on methodology from nested case-control studies, we propose a loss function that scales well to large data sets, and enables fitting of both proportional and non-proportional extensions of the Cox model. The behavior of the proposed methods is assessed in simulation studies, and the methods are shown to perform well. In particular, the non-proportional method is able to estimate a number of different forms of the survival function. A case study on customer churn is conducted, and our non-proportional method is found to perform better than the random survival forests and almost as well as binary classifiers (deep neural networks). By clustering the survival curves obtained with our non-proportional method, it is revealed that the customers may be divided in groups with quite different churn patterns. This information is not available through a binary classifier, which only gives probability estimates for fixed follow-up times.