Abstract:
|
Diagnostic classification using gene profile is an active area of research in microarray. Good experimental designs could benefit prediction accuracy through weeding out some technical bias. In this study we focus primarily on evaluating the effectiveness of blocking and randomization in classification problems. We used a unique paired datasets (Qin '14), profiled from 96 ovarian and 96 endometrial tumor samples (n*=192, p* = 210) with and without blocking and randomization. None or quantile normalization were applied before diverse classification methods were used for the evaluation. Naturally, our secondary goals are to compare the performance of those prediction methods and understand the effect of quantile normalization in conjunction with prediction methods. The evaluation was based on repeated cross-validations where prediction accuracy and reproducibility were quantified by misclassifications and feature selection frequencies respectively. In conclusion, blocking and randomization improved prediction on average (~40%), while normalization made little difference with few exceptions. Among prediction methods penalized regressions excelled across datasets. *Outcome: tumor type.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.