Dissecting the pathobiology of suspected sepsis through a comparative analysis of endothelial inflammatory and clinical prediction models

Sepsis remains a formidable challenge in critical care, and is characterized by profound circulatory and cellular abnormalities driven by both systemic inflammation and widespread endothelial dysfunction. However, the relative predictive utility of biomarkers representing these pathways versus standard clinical data is uncertain. In this analysis, we sought to conduct a comparative analysis of predictive models for forecasting two critical outcomes in sepsis patients: persistent vasopressor dependence and acute kidney injury (AKI). We prospectively enrolled a cohort of suspected sepsis patients recruited from the emergency departments of three secondary and tertiary-level teaching hospitals. We developed three distinct machine learning models via LightGBM: Model A (endothelial: angiopoietin-2, VCAM-1, and E-selectin), Model B (inflammatory: procalcitonin, CRP, and IL-6), and Model C (clinical: SOFA score and Lactate). The models were examined for their accuracy in predicting persistent vasopressor dependence and the development of KDIGO stage ≥2 AKI. For predicting persistent vasopressor dependence, the clinical model (Model C) secured a strikingly high Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.92, which was statistically superior to both the endothelial Model A (AUROC 0.53, p=0.02) and the inflammatory Model B (AUROC 0.49). For predicting AKI, the clinical model again achieved optimal results with an AUROC of 0.81, followed by the endothelial model (AUROC 0.73), although this difference was not statistically significant (p=0.38). Our findings, contrary to our initial hypothesis, demonstrate that a model based on readily available clinical data (SOFA and lactate) provides superior predictive accuracy for vasopressor dependence and AKI compared with models based on specific endothelial or inflammatory biomarker panels. This highlights the robust, integrated nature of clinical scoring systems and underscores the importance of benchmarking novel biomarker models against established clinical standards.