Adapting and evaluating deep-pseudo neural network for survival data with time-varying covariates

Area of Science:

• Biostatistics
• Machine Learning
• Survival Analysis

Background:

• Traditional Cox models may not fully capture complex survival data with time-varying covariates.
• Deep learning, specifically the Deep-pseudo survival neural network (DSNN), excels with time-invariant survival data.
• Extending DSNN to time-varying covariates offers potential for improved survival function estimation.

Purpose of the Study:

• To adapt and evaluate the Deep-pseudo survival neural network (DSNN) for survival probability prediction in the presence of time-varying covariates.
• To compare the performance of the adapted DSNN against established survival models.
• To validate the DSNN's utility in real-world survival data applications.

Main Methods:

• Adaptation of the Deep-pseudo survival neural network (DSNN) for time-varying covariates.
• Utilizing Brier scores to assess prediction accuracy at specific time points.
• Comparison with Extended Cox, Dynamic-DeepHit, and multivariate joint models on simulated data.
• Application to a real-world dataset with time-varying covariates.

Main Results:

• The adapted DSNN demonstrated strong predictive performance, with Brier scores below 0.25 for significant time-varying covariates.
• Performance was comparable to Extended Cox, Dynamic-DeepHit, and multivariate joint models on simulated data.
• The model's predictive potential was further confirmed in a real-world data application.

Conclusions:

• The adapted Deep-pseudo survival neural network (DSNN) is a viable and effective tool for survival analysis involving time-varying covariates.
• DSNN offers a competitive alternative to existing models, particularly in complex survival data scenarios.
• The study highlights the potential of deep learning for enhancing survival prediction accuracy.