A latent variable model for survival time prediction with censoring and diverse covariates

TL;DR

This paper introduces an integrative latent variable model combining factor analysis and Cox proportional hazards for improved survival time prediction in cancer, effectively handling high-dimensional, heterogeneous, and censored data.

Contribution

The novel model jointly analyzes multiple covariate types with interpretability, outperforming alternatives in simulations and revealing known subpopulations in real cancer datasets.

Findings

Model outperforms alternatives in simulations.

Competitive performance on real cancer datasets.

Visualizes known subpopulations in glioma data.

Abstract

Fulfilling the promise of precision medicine requires accurately and precisely classifying disease states. For cancer, this includes prediction of survival time from a surfeit of covariates. Such data presents an opportunity for improved prediction, but also a challenge due to high dimensionality. Furthermore, disease populations can be heterogeneous. Integrative modeling is sensible, as the underlying hypothesis is that joint analysis of multiple covariates provides greater explanatory power than separate analyses. We propose an integrative latent variable model that combines factor analysis for various data types and an exponential Cox proportional hazards model for continuous survival time with informative censoring. The factor and Cox models are connected through low-dimensional latent variables that can be interpreted and visualized to identify subpopulations. We use this model to predict survival time. We demonstrate this model's utility in simulation and on four Cancer Genome Atlas datasets: diffuse lower-grade glioma, glioblastoma multiforme, lung adenocarcinoma, and lung squamous cell carcinoma. These datasets have small sample sizes, high-dimensional diverse covariates, and high censorship rates. We compare the predictions from our model to two alternative models. Our model outperforms in simulation and is competitive on real datasets. Furthermore, the low-dimensional visualization for diffuse lower-grade glioma displays known subpopulations.