Deep Gated Neural Network With Self-Attention Mechanism for Survival Analysis

Survival analysis is commonly used to model the time distributions of the first occurrences of events of interest, and it has widespread medical applications. Many previous studies learned the relationship between risk and covariates by making strong assumptions such as proportional hazards. However, these assumptions limit the performance somewhat. Moreover, few studies consider the temporal patterns in feature effects. This paper proposed the novel framework of a deep gated neural network with self-attention mechanism (SA-DGNet) for survival analysis with single risk and competing risks. SA-DGNet transforms the problem of survival analysis into a time-series forecasting problem that treats time as an additional input covariate and estimates the probability mass function of the first hitting time. No assumptions are made about the distribution of survival times, and a deep gated neural network module is used to calculate the time-dependent and nonlinear effects of covariates on survival outcomes. Meanwhile, for enhanced data perception, a self-attention module comprising multi-scale time-aware self-attention and scaled dot-product self-attention is designed. The results of performance evaluation on multiple real-world datasets indicate that SA-DGNet significantly outperforms previous state-of-the-art methods. This study demonstrates the potential of gated neural networks and self-attention mechanisms in survival analysis, and it provides an effective method for risk prediction based on structured data.