Partially linear Cox model with neural networks for left-truncated data.

Area of Science:

• Biostatistics
• Machine Learning
• Survival Analysis

Background:

• Artificial neural networks (ANNs) excel at capturing nonlinear patterns in supervised learning.
• Left truncation in survival data, where individuals with slower disease progression are enrolled, leads to biased survival time samples.
• Traditional survival models may struggle with the complexity and bias introduced by such data.

Purpose of the Study:

• To integrate ANNs with a partially linear Cox model to handle left-truncated survival data.
• To develop a flexible model balancing parametric interpretability with the nonlinear approximation capabilities of ANNs.
• To improve the accuracy and predictive performance of survival time estimations in biased samples.

Main Methods:

• Developed a hybrid model combining parametric covariate effects with ANNs for nuisance function approximation.
• Employed conditional maximum likelihood estimation to derive a baseline hazard-free profile likelihood.
• Utilized an iterative algorithm with stochastic gradient descent for simultaneous estimation of regression parameters and ANNs.

Main Results:

• The proposed model demonstrated superior estimation accuracy compared to traditional methods in simulations.
• The method showed enhanced predictive capability on survival data.
• The integration effectively handled the bias introduced by left truncation.

Conclusions:

• The proposed ANN-integrated partially linear Cox model offers a robust solution for left-truncated survival data.
• This approach provides a valuable tool for biostatisticians and data scientists dealing with biased survival datasets.
• The method balances model interpretability with the flexibility of machine learning techniques.