High-Dimensional Variable Selection With Competing Events Using Cooperative Penalized Regression.

Area of Science:

• Biostatistics
• Bioinformatics
• Computational Biology

Background:

• Variable selection is crucial for high-dimensional data analysis, especially for survival outcomes with competing risks.
• Existing methods like cause-specific penalized Cox regression often ignore potentially shared information between competing events.

Purpose of the Study:

• To adapt the feature-weighted elastic net (fwelnet) for survival outcomes and competing risks.
• To develop a novel 'cooperative penalized regression' method that accounts for shared effects between competing risks.

Main Methods:

• Proposed an algorithm that fits two alternating cause-specific models, incorporating prior information from the complementary model.
• Implemented a cooperative penalized regression approach where coefficients from one model inform the penalization weights of the other.
• Evaluated variable selection performance using positive predictive value and false positive rate on simulated and real-world (bladder cancer) data.

Main Results:

• The cooperative penalized regression method demonstrated superior performance in selecting informative features and excluding uninformative ones compared to benchmarks.
• Performance was comparable to cause-specific penalized Cox regression in scenarios lacking shared effects between risks.
• Validated on genomics and bladder cancer microarray data, showcasing practical applicability.

Conclusions:

• Cooperative penalized regression effectively models competing risk data by utilizing shared information between cause-specific models.
• The method offers improved variable selection accuracy for high-dimensional survival data with competing risks.
• This approach enhances the analysis of complex survival data by integrating complementary event information.