Comparison of Variable Selection Methods for Time-to-Event Data in High-Dimensional Settings.

Area of Science:

• Oncology
• Bioinformatics
• Statistical Genetics

Background:

• Molecular signatures are vital for personalized oncology, enabling targeted therapies.
• Concerns exist regarding the statistical validity and biological relevance of current signature development methods.
• High-dimensional genomic data presents unique challenges for accurate predictor selection.

Purpose of the Study:

• To evaluate and compare six statistical selection methods for molecular signatures in high-dimensional settings with survival endpoints.
• To assess the stability and prognostic performance of these methods using simulated and real-world data.
• To identify optimal methods for developing reliable molecular signatures in oncology.

Main Methods:

• Investigated six selection methods: LASSO, its extensions, component-wise boosting, and random survival forests (RSF).
• Employed a resampling algorithm with data splitting on nine simulated high-dimensional datasets.
• Evaluated selection stability on training sets, predictor intersection, and prognostic performance on validation sets.

Main Results:

• All tested methods exhibited a high false discovery rate (FDR), with poor overlap between selected predictors.
• RSF selected a large number of variables, reducing its efficiency on validation sets.
• LASSO-pcvl demonstrated superior performance in high-dimensional settings, reducing gene count and minimizing FDR, but still yielded false positives.

Conclusions:

• Selection stability for molecular signature predictors is generally poor in high-dimensional genomic data, but improves with larger training datasets.
• LASSO-pcvl is recommended for very high-dimensional scenarios due to its balance of gene reduction and FDR control.
• Further methodological development is required to address false positives, and multidisciplinary collaboration is essential for robust molecular signature development.