Subtype classification and heterogeneous prognosis model construction in precision medicine.

Area of Science:

• Biostatistics
• Genomics
• Cancer Research

Background:

• Cancer and other common diseases exhibit heterogeneity, necessitating subtype discovery and identification of shared/unique risk factors.
• High-throughput technologies offer rich data for this goal, but require advanced statistical methods.
• Existing methods for heterogeneity and variable selection are limited, especially in semiparametric survival analysis.

Purpose of the Study:

• To develop a variable selection method for the finite-mixture Cox model to address heterogeneity in survival analysis.
• To identify subtype-specific risk factors from high-dimensional predictors.
• To improve survival prediction by incorporating disease heterogeneity.

Main Methods:

• Proposed a variable selection method using regularization regression for the finite-mixture Cox model.
• Developed an expectation-maximization algorithm for numerical computation.
• Validated the method through simulations and application to an ovarian cancer gene expression dataset.

Main Results:

• The proposed method effectively reveals disease heterogeneity and selects important, subtype-specific risk factors.
• Estimators demonstrate oracle properties with appropriate penalty parameter selection.
• The method consistently improved survival probability prediction in both training and test datasets for ovarian cancer.

Conclusions:

• The developed regularization method successfully integrates variable selection into semiparametric finite-mixture Cox models.
• This approach enhances the understanding of disease subtypes and their associated risk factors.
• The method offers a robust tool for improving prognostic models in heterogeneous diseases like cancer.