Variable Selection via Fused Sparse-Group Lasso Penalized Multi-state Models Incorporating Molecular Data.

Area of Science:

• Statistics
• Biostatistics
• Computational Biology

Background:

• High-dimensional data in multi-state models necessitates parsimonious modeling strategies.
• Linking covariate effects across transitions is crucial for joint variable selection.
• Reducing model complexity via homogeneous covariate effects is a key challenge.

Purpose of the Study:

• To develop a data-driven variable selection method for multi-state models.
• To propose the fused sparse-group lasso (FSGL) for parsimonious model building.
• To integrate homogeneous covariate effects across transitions using regularization.

Main Methods:

• Utilized Cox-type regression within a multi-state model framework.
• Developed the fused sparse-group lasso (FSGL) penalty combining difference and group penalization.
• Adapted the alternating direction method of multipliers (ADMM) for optimization.
• Evaluated the method using simulation studies and acute myeloid leukemia (AML) data.

Main Results:

• The FSGL method successfully selects sparse models with relevant transition-specific and cross-transition effects.
• Demonstrated the benefit of the combined penalty over global lasso regularization.
• The ADMM algorithm efficiently handles transition-specific hazards regression.

Conclusions:

• The proposed FSGL approach offers an effective strategy for variable selection in high-dimensional multi-state models.
• This method aids in identifying parsimonious models by leveraging shared covariate effects.
• The application to AML data highlights the practical utility of the developed technique.