Without consistent rankings from the predictors, good classification is not possible.
Only 25% of those with the "bad" classification were randomized to Arm B while 62% of those with the "good" classification were assigned to Arm B and so in this small study, the treatment arm assignment might confound the assessment of the serum proteomic classifier and vice-versa.
Good classifications were observed in the PCA score plots for days 1 and 3 (n=10) (Fig. 6A).
The relative PFS and OS advantage of VS Good over VS Poor classification was similar in both treatment arms, with no significant interaction, suggesting that VeriStrat in this setting may not be predictive of response to erlotinib but rather prognostic in patients with pancreatic cancer 23.
Good classification performance was observed when replicating results with PCR data obtained from the same cohort of patients used in the model derivation phase.
In brief, for the first training data that contains only histone acetylation information, good classification accuracy was achieved for around one third of all TFs tested (e.g. Table S1a); however, for the second dataset that includes both histone methylation and histone acetlylation features, almost half of all TFs tested were well classified by histone modifications (e.g. Table 1).
This phenomenological classification was good enough to predict new particles which could be found in the experiments.
Therefore it is recommended that performance be compared for signatures involving 1, 2,3, 5 and 10 genes to determine the signature length such that good classification performance is obtained and further gains in classification performance with longer signatures are minor [ 33].
If a label for its object category is used for each image during training, this may help to produce good classification, but is very subject to over-fitting, in which small pixel changes in an image that do not affect which object it is interpreted as by humans may lead to it being misclassified (Krizhevsky et al. 2012; Szegedy et al. 2014).
