Recent research has unravelled the clinical potential of profiling the blood transcriptome to diagnose diseases. However, resulting molecular marker sets comprised features with varying robustness and performance, depending on the dimension of training data. Thus, we investigated patterns that are inherent in large-scale data and suitable for feature selection in application to blood samples from septic patients. By integrating >300 microarray samples in correlation and enrichment analysis, we found general response patterns including a vast majority of co-expressed genes. Differentially expressed genes significantly mapped to immune response-associated categories and revealed strongly correlating upregulated genes related to antimicrobial functions. Classifiers using >20 uncorrelated features from enriched functional categories performed with 85% correct classification on average (10-fold cross-validation), comparable with correlated features, whilst single genes achieved up to 83% correct classifications in identifying septic patients. Independent interplatform comparison, however, validated only a subset of these features, including the antimicrobial cluster (area under the receiver operating characteristic curve >0.8). Based on these results, we propose feature selection for classification incorporating correlation and enriched functional categories to obtain robust marker candidates. Results of this transcriptomic meta-analysis suggest age-independent diagnostic opportunities, although further observational and animal interventional experiments are required to confirm the relevance of antimicrobial genes in sepsis.