Combined Use of Three Machine Learning Modeling Methods to Develop a Ten-Gene Signature for the Diagnosis of Ventilator-Associated Pneumonia.

BACKGROUND This study aimed to use three modeling methods, logistic regression analysis, random forest analysis, and fully-connected neural network analysis, to develop a diagnostic gene signature for the diagnosis of ventilator-associated pneumonia (VAP). MATERIAL AND METHODS GSE30385 from the Gene Expression Omnibus (GEO) database identified differentially expressed genes (DEGs) associated with patients with VAP. Gene Ontology (GO) and the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment identified the molecular functions of the DEGs. The least absolute shrinkage and selection operator (LASSO) regression analysis algorithm was used to select key genes. Three modeling methods, including logistic regression analysis, random forest analysis, and fully-connected neural network analysis, also known as also known as the feed-forward multi-layer perceptron (MLP), were used to identify the diagnostic gene signature for patients with VAP. RESULTS Sixty-six DEGs were identified for patients who had VAP (VAP+) and who did not have VAP (VAP-). Ten essential or feature genes were identified. Upregulated genes included matrix metallopeptidase 8 (MMP8), arginase 1 (ARG1), haptoglobin (HP), interleukin 18 receptor 1 (IL18R1), and NLR family apoptosis inhibitory protein (NAIP). Down-regulated genes included complement factor D (CFD), pleckstrin homology-like domain family A member 2 (PHLDA2), plasminogen activator, urokinase (PLAU), laminin subunit beta 3 (LAMB3), and dual-specificity phosphatase 2 (DUSP2). Logistic regression, random forest, and MLP analysis showed receiver operating characteristic (ROC) curve area under the curve (AUC) values of 0.85, 0.86, and 0.87, respectively. CONCLUSIONS Logistic regression analysis, random forest analysis, and MLP analysis identified a ten-gene signature for the diagnosis of VAP.