DeepSurv landmarking: a deep learning approach for dynamic survival analysis with longitudinal data

Abstract

Risk prediction models quantify the probability of an individual experiencing specified events, there has significant attention to utilize longitudinal data during follow-up to dynamically update risk estimates. This paper devotes to a dynamic prediction model that combines survival analysis and deep learning techniques, which integrates the deep learning network-DeepSurv into the landmarking framework, to predict time-to-event outcomes with longitudinal covariates. Precisely, at each landmark time point, new information of the dataset is reconstructed, and the DeepSurv network is employed to capture the non-linear relationship between covariates and hazard rates, thereby give the dynamic prediction of survival functions. Furthermore, the importance of different features at each landmark time point is evaluated and ranked to estimate the impact of covariates on predictive outcomes over time. Performances on simulation experiments demonstrate that the proposed DeepSurv landmarking model outperforms state-of-art landmark methods, especially when the dependencies between covariates and the survival process are unknown. Additionally, the model exhibits remarkable dynamic prediction capabilities in the real heart-valve-transplantation-surgery data.