Case-Base Neural Network: Survival analysis with time-varying, higher-order interactions.

Area of Science:

• Biostatistics
• Machine Learning
• Computational Biology

Background:

• Existing neural network survival models struggle with complex time-varying interactions and baseline hazards.
• Regression-based methods may lack the flexibility to capture intricate covariate effects.

Purpose of the Study:

• Introduce Case-Base Neural Networks (CBNNs) for enhanced survival analysis.
• Develop a flexible deep learning framework capable of modeling complex time-varying effects and baseline hazards.

Main Methods:

• Proposed Case-Base Neural Networks (CBNNs) integrating case-base sampling with neural networks.
• Utilized a novel sampling scheme and data augmentation to handle censored data.
• Constructed a feed-forward neural network with time as an input to predict event probabilities and estimate hazard functions.

Main Results:

• CBNNs demonstrated superior performance over regression and other neural network survival models in a simulation study with complex baseline hazards and time-varying interactions.
• In three real-world case studies, CBNNs outperformed competing models in two applications and showed comparable performance in the third.

Conclusions:

• Combining case-base sampling with deep learning provides a flexible and effective framework for survival analysis.
• CBNNs successfully model time-varying effects and complex baseline hazards in single event survival data.
• The proposed method offers a data-driven approach with improved predictive performance for survival outcomes.