Accounting for grouped predictor variables or pathways in high-dimensional penalized Cox regression models

Area of Science:

• Biostatistics
• Genomics
• Machine Learning

Background:

• Standard lasso methods for variable selection in high-dimensional time-to-event data do not account for predictor grouping.
• These methods can also exhibit high false discovery rates, particularly with complex biological data like genomics.
• Functional groups in predictors, such as biological pathways, require specialized selection approaches.

Purpose of the Study:

• To develop a regularized regression model that effectively selects grouped variables in high-dimensional data.
• To address the limitations of standard lasso, including its handling of grouped predictors and high false discovery rates.
• To propose a novel weighting strategy for adaptive lasso that leverages both individual variable and group-level information.

Main Methods:

• Evaluated various penalizations within a Cox model framework for selecting grouped variables.
• Proposed diverse weights for the adaptive lasso method, including the novel Max Single Wald by Single Wald (MSW*SW) weighting.
• Compared the proposed method against standard lasso, cMCP, gel, IPF-Lasso, and SGL using simulations and real-world gene expression data from breast cancer patients.

Main Results:

• The adaptive lasso with MSW*SW weighting demonstrated superior variable selection capabilities compared to existing methods.
• The proposed method effectively utilized both individual variable effects and group-level information for enhanced prediction.
• Simulations confirmed the method's ability to manage the trade-off between false discovery and false negative rates.

Conclusions:

• The adaptive lasso with MSW*SW weighting successfully integrates grouping structure and individual variable information.
• This approach significantly reduces the false discovery rate compared to competing methods.
• The method offers improved variable selection for high-dimensional data with inherent grouping, as illustrated in breast cancer genomics.