Decentralized Learning Framework of Meta-Survival Analysis for Developing Robust Prognostic Signatures

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Developing reliable gene expression-based prognostic models is challenging due to limited sample sizes, leading to overfitting and false discoveries.
• Combining data from multiple studies can increase statistical power but is hindered by biological heterogeneity across datasets.
• There is a need for advanced meta-survival analysis methods to overcome these limitations.

Purpose of the Study:

• To present a novel decentralized learning framework for meta-survival analysis that does not require data aggregation.
• To alleviate the influence of data heterogeneity and improve survival prediction performance.
• To develop robust multigene prognostic signatures from diverse gene expression datasets.

Main Methods:

• Transformed gene expression profiles into normalized percentile ranks for platform-agnostic features.
• Employed Stouffer's meta-z approach and Harrell's concordance index for gene selection.
• Utilized survival discordance as a scale-independent loss function, optimizing the model by minimizing a joint loss function across individual datasets.

Main Results:

• The proposed decentralized learning method outperformed single prognostic genes, single-dataset signatures, merged-dataset signatures, and existing meta-analysis methods.
• Demonstrated superior performance on 31 public microarray datasets comprising 6,724 cancer samples.
• Outperformed established clinically applicable multigene signatures.

Conclusions:

• The decentralized learning approach effectively performs meta-analysis of gene expression data.
• This framework enables the development of robust and accurate multigene prognostic signatures.
• The method offers a powerful solution for leveraging multi-study gene expression data while managing heterogeneity.