Identification of potential biological targets in sepsis based on random forest and artificial neural network models

BACKGROUND: Machine learning and artificial neural networks are increasingly making new research progress in various human diseases. OBJECTIVE: To identify key gene markers related to sepsis based on random forest (RF) and artificial neural network (ANN) models. METHODOLOGY: Key genes (KG) were screened through RF based on GSE9960, GSE13904, and GSE26440 datasets, and the diagnostic model of core genes was constructed using ANN analysis, which was validated. RESULTS: GSE9960, GSE13904, and GSE26440 were merged as the training set, and 452 differentially expressed genes (DEGs) were identified. Then, GO and KEGG enrichment analysis, as well as Cytoscape module analysis, were used to explore the mechanisms related to sepsis. Subsequently, these DEGs were further screened through RF to obtain key genes UPP1, ZNF600, GPR160, TRAT1, C3AR1, NAIP, MAL, and F5. An ANN method was used to construct a screening model. The AUC metrics obtained from the training and validation cohorts were quantitatively assessed through ROC curve analysis, yielding discriminative performance values of 0.969 and 0.792, respectively. The CIBERSORT-based immune profiling revealed a marked depletion of lymphocyte populations (including B cells and T cells) in the experimental cohort than healthy controls, with regulatory T cell subsets demonstrating a paradoxical increase in abundance. Sepsis immune suppression was the dominant process. CONCLUSIONS: The core-gene model constructed by the RF and ANN method is a robust method for predicting, diagnosing, or treating sepsis, with nice predictive accuracy. It may be used for sepsis-related medical decisions and provide new entry routes for the treatment or prognosis improvement of sepsis.