Average results of classification performed on the Training Set.

<p>The signature obtained by just selecting features that best complement Gómez Ravetti and Moscato's <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0017481#pone.0017481-GmezRavetti1" target="_blank">[3]</a>, S1, improves Ray et al.'s <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0017481#pone.0017481-Ray1" target="_blank">[1]</a> classification accuracy on the training set of 93% to 94.4%, with improvements both in sensitivity and specificity. This signature includes the following features and meta-features: EGF, IL-1, IL-3, TNF-, G-CSF, “BLC-RANTES”, “MIP-1d-IL-11”, “TNF- - ANG-2”, “TNF- - FAS” and “IL-11-I-TAC”. The performance of the second signature, S2, obtained in the same manner after discarding samples s3, s7, s47, s66 and s77 is still better than that of the other signatures, with an even greater gap, reaching an average of 98% of learning accuracy. Even though this does not necessarily imply an improvement also on independent test sets, this provides good evidence that the discarded samples were indeed problematic. This signature includes the following features and meta-features: EGF, IL-1, TNF-, G-CSF, “EGF-IGFBP-2”, “GM-CSF-IL-1”, “IL-1-IL-11”, “MIP-1d-NT-3”, “PDGF-BB-VEGF-B” and “TNF--ANG-2”. The signature obtained by just discarding the single features from S2, S3, also shows a very good performance on the training set. It is remarkable that the average results differed by less than 1% from those from the 4 feature and 6 meta-feature signature. That suggests that the single features were not playing a key role to distinguish between AD and NDC, and supports the theory that there is useful information provided by the meta-features to distinguish between the classes.