DyVAESurv: A VAE‐Enhanced Model for Dynamic Survival Analysis

Survival analysis plays a crucial role in discovering the relationship between covariates and event times, which is widely applied in individual risk prediction. Traditional methods often rely on static features, resulting in the difficulty of capturing time dependencies, while dynamic survival models have limitations in capturing heterogeneity among individuals. To address these issues, a deep learning model, integrating original features with latent representations generated by a variational autoencoder (VAE), is proposed to better characterize individual survival trajectories. The model employs a dynamic feature extractor that selects the appropriate Transformer or LSTM branch based on the sequence length, thus adapting to different time-series data. The model retains the information of the original data upon joint representation, it also utilizes the latent variables of the VAE to effectively capture the heterogeneity among individuals. Experimental results demonstrate that the proposed model achieves superior performance in terms of C-index and IBS, providing a novel solution for dynamic prediction based on survival analysis.