Predictive modeling of ARDS mortality integrating biomarker/cytokine, clinical and metabolomic data

Abstract

Acute Respiratory Distress Syndrome (ARDS), characterized by the rapid onset of respiratory failure and mortality rates of ∼40%, remains a significant challenge in critical care medicine. Despite advances in supportive care, accurate prediction of ARDS mortality remains challenging, resulting in delayed delivery of targeted interventions and effective disease management. Traditional critical illness severity scores lack specificity for ARDS, underscoring the need for more precise prognostic tools for ARDS mortality. To address this crucial gap, we employed a multimodal approach to predict ARDS patients utilizing a comprehensive dataset comprised of integrated clinical, metabolomic, and biochemical/cytokine data from ARDS patients (collected within hours of ICU admission) to develop and validate predictive models of ARDS mortality risk. The most robust multimodal data model generated demonstrated superior predictive capability with an area under the curve (AUC) of 0.868 on the test set and 0.959 on the validation set. Notably, this model achieved perfect specificity in identifying non-survivors in the validation cohort, highlighting potential utility in guiding early and targeted interventions in ICU settings. Metabolomic analysis revealed significant alterations in crucial pathways associated with ARDS mortality with tryptophan metabolism, particularly the kynurenine pathway, emerging as the most significantly enriched metabolic route, as well as the NAD+ metabolism/nicotinamide phosphoribosyltransferase (NAMPT) and glycosaminoglycan biosynthesis pathways. These metabolic derangements were strongly confirmed by lipidomic/metabolomic analysis of lung tissues from a porcine sepsis/ARDS model. Together, these findings demonstrate the promise of integrating multimodal data to improve ARDS prognostication and to provide important insights into the complex metabolic derangements underlying severe ARDS. Identification of metabolic signatures, such as kynurenine and NAD+ metabolism/NAMPT pathways, may serve as a foundation for developing personalized and effective targeted interventions and management strategies for ARDS patients.

Keywords: ARDS; ICU; Machine learning; Mortality prediction; NAMPT; Omics.