A comparative study of survival models for breast cancer prognostication based on microarray data: does a single gene...

Area of Science:

• Genomics
• Bioinformatics
• Cancer Research

Background:

• Breast cancer (BC) prognostication is complex due to molecular heterogeneity in tumors.
• Gene expression profiling offers insights into BC biology and can improve prognostication.
• Existing methods for BC survival prediction require quantitative accuracy assessment.

Purpose of the Study:

• To quantitatively assess the predictive accuracy of state-of-the-art data analysis techniques for BC microarray data.
• To evaluate prediction accuracy using an independent and thorough framework.
• To compare the performance of complex data mining tools against simpler models.

Main Methods:

• Utilized a comprehensive framework to evaluate BC microarray data.
• Applied various sophisticated data mining and analysis techniques.
• Compared prediction accuracy across different methods, including univariate and complex models.

Main Results:

• Complex prediction methods showed no significant improvement over a simple univariate model based on a single proliferation gene.
• High dimensionality, limited samples, and noise in microarray data challenge complex prediction methods.
• Proliferation appears to be a highly relevant biological process for BC prognostication.

Conclusions:

• Overly complex methods may not enhance BC prediction accuracy sufficiently to justify their loss of interpretability.
• Simpler models focusing on key biological processes like proliferation may be more effective for BC prognostication.
• The R package 'survcomp' is available for comparative analysis of survival prediction methods.