Deep Learning for Survival Analysis: A Review

Survival analysis (SA), or equivalently time-to-event analysis, comprises a set of techniques enabling the unbiased estimation of the distribution of outcome variables that are partially censored, truncated, or both. Usually, the outcome is given by the time until the occurrence of an event such as death, system failure, or time to remission. Non-parametric methods like the Kaplan-Meier estimator [1] are baseline tools still used today, yet semi-parametric methods received the most attention historically, in particular the Cox proportional hazards regression model [2] and its extensions. Since the early 2000s, Machine Learning (ML) methods have been successfully adapted to survival tasks: e.g., Random Survival Forest [3] and boosting-based methods [4]. These methods often outperform traditional statistical models in terms of predictive power [5] (see [6] and [7] for detailed discussions). Neural networks (NNs) had already been applied to survival tasks in the 1990s [8, 9], but were shallow and restricted to the most standard survival settings.