Disentangling Whether from When in a Neural Mixture Cure Model for Failure Time Data

Area of Science:

• Biostatistics
• Machine Learning
• Causal Inference

Background:

• Mixture cure models are essential for analyzing data with a non-susceptible population subgroup.
• Traditional models often rely on restrictive distributional assumptions.
• Selection biases can impact the estimation of failure and censoring densities.

Purpose of the Study:

• To develop a flexible, neural network-based mixture cure model.
• To improve the prediction of failure timing and probability.
• To address biases in observational failure and censoring time data.

Main Methods:

• Utilized representation learning and causal inference principles.
• Developed a neural network architecture for the mixture cure model.
• The model is free of distributional assumptions.

Main Results:

• Achieved improved prediction accuracy for failure timing.
• Successfully disentangled information on failure probability from failure timing.
• Mitigated the effects of selection bias on density estimations.

Conclusions:

• The proposed neural network mixture cure model offers enhanced flexibility and accuracy.
• This approach can differentiate predictors of failure occurrence versus timing.
• It provides a robust method for analyzing biased observational datasets.