Machine Learning Identification of Transcriptional Signatures for Predicting Sepsis-Associated Acute Respiratory Distress Syndrome

Le résumé fourni par la source

Abstract Rationale: Sepsis-associated acute respiratory distress syndrome (ARDS) is a life-threatening condition that significantly increases mortality among critically ill patients and there are limited effective treatment options available. Identifying key molecular pathways involved in ARDS could drive targeted therapy development and inform clinical trials. This study applied advanced machine learning (ML) techniques to analyze transcriptomic data from septic patients, aiming to uncover differentially expressed genes (DEGs) that predict the development of ARDS and provide insights into its underlying molecular mechanisms. Methods: We analyzed transcriptomic data from three publicly available datasets (GSE10474, GSE32707, GSE66890) comprising whole blood samples from 135 patients (77 with sepsis alone and 58 with sepsis-associated ARDS) collected within 48 hours of ICU admission. Data were harmonized and batch effects corrected to ensure comparability. Two ML algorithms were employed: Empirical Bayesian Elastic Net (EBEN), a generalized linear model, and eXtreme Gradient Boosting (XGBoost), a gradient boosting method. The models were trained to distinguish septic patients with and without ARDS, and their performance was assessed using the area under the receiver operating characteristic curve (AUROC). A 10-fold cross-validation approach was employed to construct the model, with 109 patients in the training set, and 26 patients in the test set. A multivariable logistic regression model consisting of the DEGs identified by both models was developed and evaluated using the same dataset. Results: The EBEN model identified 11 DEGs, achieving an AUROC of 0.883 (95% CI: 0.819-0.948) in the training set and 0.909 (95% CI: 0.792-1.000) in the test set. The XGBoost model identified 33 DEGs, demonstrating superior predictive accuracy with an AUROC of 0.996 (95% CI: 0.990-1.000) in the training set and 0.964 (95% CI: 0.898-1.000) in the test set. Seven overlapping DEGs were identified by both models, including TREM1, FGL2, and FTH1, which are involved in inflammatory pathways relevant to ARDS pathogenesis. The AUROC of the seven DEG model was 0.897 (95% CI:0.748-1.000). Conclusions: Our study demonstrates that the XGBoost model provides superior predictive accuracy for sepsis-associated ARDS, significantly outperforming both the EBEN model and previous ML approaches. The identified DEGs, particularly those consistently highlighted across models, reveal novel molecular pathways that could be therapeutically targeted. These DEGs include immune-related genes that may play pivotal roles in disease progression, providing a deeper understanding of the molecular drivers of ARDS. Integrating these biomarkers into clinical workflows (especially if validated with proteomics studies) could enhance early identification and treatment stratification.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.