Title: Deep learning for time to event data
Authors: Federico Ambrogi - University of Milan (Italy) [presenting]
Thomas Scheike - Section of Biostatistics University of Copenhagen (Denmark)
Abstract: The use of neural networks for developing prediction models with survival data has received much attention in the literature. The developments with machine learning methods and with computational facilities have renewed interest in such kind of models and with the use of deep neural networks. There are some proposals already available for adapting deep learning regression models to time to event data. We propose a simple method, easily generalisable to complex settings, such as competing risks or semi-competing risks, working with standard software for deep neural networks in R and Python, namely Keras. The presented approach is based on the use of standard binomial estimating equations with weights. A similar approach is the one based on pseudo values allowing to use standard software provided it is possible to have an appropriate link function. An application based on SEER data is presented with a comparison of model calibration and prediction error with respect to standard methods.