Model-based optimization of subgroup weights for survival analysis.

Area of Science:

• Biostatistics
• Bioinformatics
• Computational Biology

Background:

• Reliable prediction models for specific cancer subgroups are difficult to develop due to limited sample sizes and high censoring rates in survival analysis.
• Pooling data from different patient subgroups can reduce variance but increase bias due to cohort heterogeneity.

Purpose of the Study:

• To propose a subgroup-based weighted likelihood approach for survival prediction using high-dimensional genetic covariates.
• To improve the accuracy and reliability of cancer subgroup survival prediction models by leveraging data from similar cohorts.

Main Methods:

• A subgroup-based weighted likelihood approach is proposed for survival prediction.
• Model-based optimization (MBO) is used to identify optimal models by determining weights for incorporating data from other subgroups.
• The method is applied to Cox models for survival prediction with high-dimensional genetic data.

Main Results:

• The proposed approach yields prediction models with competitive quality for specific cancer subgroups.
• The identified weights effectively reflect the similarity between cancer subgroups, with values ranging from 0 to 1.
• The method was evaluated on lung cancer cohorts with gene expression data.

Conclusions:

• The subgroup-based weighted likelihood approach offers a promising strategy to overcome limitations of small sample sizes and cohort heterogeneity in cancer survival prediction.
• This method allows for the development of more accurate and robust prediction models by selectively incorporating data from similar patient subgroups.
• The freely available R-package 'mlrMBO' facilitates the implementation of this approach.