Biomedical Informatics for Cancer Research by Michael F. Ochs John T. Casagrande & Ramana V. V. Davuluri

Author:Michael F. Ochs, John T. Casagrande & Ramana V. V. Davuluri
Language: eng
Format: epub
Publisher: Springer US, Boston, MA

We see 15/60 misclassifications, probably close enough given the unweighted approach. This is likely an underestimate of the true misclassification rate because the role of the given data in driving the choice of the signature is not factored in.

There are three features of Michiels’ algorithm, which attempts explicitly to address the uncertainty connected with data-driven signature selection, that depart from the analysis just shown. First, Michiels uses centroid correlation as opposed to k-nearest neighbors. Second, Michiels forces a signature size of 50 genes. Third, Michiels requires that the training set be balanced on classes to be predicted and that the test set include at least one element of each class. These constraints are incompatible with leave-one-out cross-validation.

It is easy to show that centroid correlation combined with LOO cross-validation yields a misclassification rate for the eight-gene signature that is considerably below 50%:

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.