C-mix: A high-dimensional mixture model for censored durations, with applications to genetic data

Area of Science:

• Biostatistics
• Computational Biology
• Genomics

Background:

• Accurate prognosis and risk stratification are crucial for personalized cancer medicine.
• Existing survival models struggle with high-dimensional biomedical data, limiting their clinical utility.

Purpose of the Study:

• To introduce a novel supervised learning mixture model (C-mix) for censored duration data.
• To enable simultaneous subgroup detection and risk ordering of cancer patients using high-dimensional covariates.
• To improve survival prediction accuracy in complex datasets.

Main Methods:

• Developed the C-mix model incorporating Elastic-Net penalization for feature selection in high-dimensional settings.
• Employed an efficient Quasi-Newton Expectation Maximization algorithm for model inference.
• Validated performance through extensive Monte Carlo simulations and analysis of public cancer genomics datasets.

Main Results:

• The C-mix model effectively identifies patient subgroups with distinct prognoses.
• Elastic-Net penalization successfully pinpointed relevant covariates for survival prediction.
• C-mix demonstrated superior performance compared to penalized CURE and Cox models, evidenced by higher C-index and AUC(t) values.

Conclusions:

• The C-mix model offers a powerful and accurate approach for personalized medicine in oncology.
• It addresses the challenge of high-dimensional data in survival analysis.
• This method advances risk stratification and prognosis prediction for cancer patients.