Skip to main content

Robust gene selection and committee of single gene classifiers for breast cancer survival prediction

Friday, November 08, 2013 — Poster Session IV

2:00 p.m. – 4:00 p.m.

FAES Academic Center (Upper-Level Terrace)




  • H. H Yang
  • M. P. Lee


Gene expression signatures have been widely used to predict cancer survival. However, predictors developed on the training samples often failed on external validation samples. The grand challenge is to improve the generalization of such predictors. Thus, we developed an analytic method to address the issue of robustness with feature selection and feature combination. We selected features/genes that had consistent prediction in the bootstrap samples. Considering each selected gene as a weak predictor, and then by the crowd voting method we built a strong committee predictor using majority rule. Our method had good generalization on the validation set. We analyzed n=997 breast tumor samples in the training set and found that 18 genes could predict survival consistently in the bootstrap samples. The committee predictor was applied to n=995 validation samples. We applied the Cox Proportional Hazards Model to evaluating the performance of the gene committee on validation samples. After adjusting for age, ER, tumor size, and node (which were all significant in the model), the committee predictor showed a strong independent effort with a hazard ratio HR=1.39 and p-value=0.001.

back to top