Conceitos Básicos
The proposed Variational Deep Survival Machine (VDSM) models the conditional survival function as a mixture of individual parametric survival distributions, while leveraging Variational Autoencoders (VAEs) to learn better clustering representations of the input covariates, leading to improved long-term survival time predictions.
Resumo
The paper presents two novel models, VDSM-cat and VDSM-clus, which combine the Deep Survival Machine (DSM) approach with Variational Autoencoders (VAEs) to improve survival time prediction in the presence of censored data.
Key highlights:
- DSM estimates the conditional survival function as a mixture of individual parametric survival distributions, without strong assumptions of proportional hazards.
- VDSM-cat introduces a Categorical VAE to generate the latent variables for clustering the input covariates.
- VDSM-clus uses a Generative Clustering VAE to learn a Gaussian Mixture Model representation of the latent variables.
- The VDSM models are trained end-to-end by jointly optimizing the VAE loss and the regression loss.
- Experiments on the SUPPORT and FLCHAIN datasets show that the VDSM models can achieve superior long-term survival time predictions compared to the original DSM.
- The improved performance is attributed to the VAE's ability to learn better clustering representations of the input data, which helps the final survival time prediction.
Estatísticas
The SUPPORT dataset consists of 9,105 terminally ill patients on life support, with a median survival time of 58 days.
The FLCHAIN dataset includes 6,524 individuals with covariates such as age, gender, serum creatinine, and presence of monoclonal gammapothy.
Citações
"Our goal is to learn a better cluster assignment with VAE models."
"We demonstrate the superior result of our model prediction in the long-term."