The Lq- NORM LEARNING FOR ULTRAHIGH-DIMENSIONAL SURVIVAL DATA: AN INTEGRATIVE FRAMEWORK

Area of Science:

• Biostatistics
• Genomics
• Computational Biology

Background:

• Precision medicine generates high-dimensional survival outcome data with numerous predictors.
• Overfitting and low predictability are common challenges in models with excessive covariates.
• Existing variable screening methods often rely on specific modeling assumptions.

Purpose of the Study:

• To propose a novel, model-free L-norm learning procedure for variable screening.
• To develop an integrative framework for identifying predictors impacting censored survival outcomes.
• To address the issue of false negatives in high-throughput data analysis.

Main Methods:

• A model-free L-norm learning procedure is introduced.
• The framework incorporates Cramér-von Mises and Kolmogorov criteria as special cases.
• A scheme combining results from different q values is proposed to reduce false negatives.

Main Results:

• The proposed method demonstrates sure screening properties.
• Simulation studies confirm the utility of the method.
• The approach was successfully applied to a multiple myeloma patient survival study.

Conclusions:

• The L-norm learning procedure offers a robust, model-free approach to variable screening.
• The integrative framework effectively detects predictors with varying impact levels on survival.
• The method enhances the identification of relevant genetic signatures for patient outcomes.