High-dimensional variable selection and prediction under competing risks with application to SEER-Medicare linked...

Area of Science:

• Biostatistics
• Machine Learning
• Survival Analysis

Background:

• Competing risk analysis is crucial for understanding multiple event types.
• Existing regression models include cause-specific hazards and subdistribution hazards models.
• High-dimensional predictors pose challenges for these models.

Purpose of the Study:

• To explore the accuracy of prediction and variable selection using machine learning under competing risk models.
• To investigate the performance of these methods with high-dimensional data.
• To identify optimal approaches for analyzing SEER-Medicare data.

Main Methods:

• Simulation experiments were conducted to assess prediction accuracy and variable selection.
• Machine learning methods were evaluated under both cause-specific and subdistribution hazards models.
• Various penalty parameter selection strategies were explored.

Main Results:

• The study compared the performance of different machine learning approaches for competing risk analysis.
• Optimal methods were identified based on simulation results.
• The findings inform the application of these methods to real-world datasets.

Conclusions:

• Machine learning methods show promise for competing risk analysis with high-dimensional predictors.
• The study provides guidance on selecting appropriate methods and parameters.
• The findings are applicable to predicting mortality in prostate cancer patients.