A DNN-Based Weighted Partial Likelihood for Partially Linear Subdistribution Hazard Model.

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Deep learning models have shown success in statistical learning.
• Existing deep partially linear models are limited to standard survival analysis.
• Competing risks scenarios require specialized statistical approaches.

Purpose of the Study:

• To extend deep partially linear models to the complex field of competing risks analysis.
• To propose the deep partially linear subdistribution hazard model (DPLSHM).
• To develop a robust method for evaluating model predictive performance in competing risks settings.

Main Methods:

• Development of the deep partially linear subdistribution hazard model (DPLSHM).
• Introduction of a time-dependent Area Under the Curve (AUC) method for competing risks data.
• Theoretical analysis including asymptotic normality and convergence rates for model components and AUC estimates.

Main Results:

• The DPLSHM demonstrates excellent estimation and prediction capabilities.
• The proposed time-dependent AUC method provides reliable performance evaluation.
• Theoretical analysis confirms the model's asymptotic normality and optimal convergence rates.

Conclusions:

• The DPLSHM is a powerful new tool for analyzing competing risks data.
• The developed AUC method enhances the evaluation of predictive accuracy in complex survival scenarios.
• The model shows strong performance in both simulations and real-world applications.