Diagnosing Sepsis Through Proteomic Insights: Findings from a Prospective ICU Cohort

Introduction

Sepsis diagnosis remains clinical and heterogeneous. We hypothesized that a proteomics-informed machine-learning approach could identify a small, easy-to-use, and optimized set of clinical variables to complement or potentially outperform SOFA.

Methods

We conducted a prospective, single-center, observational study in an academic intensive care unit. Plasma from critically ill patients with and without sepsis was analyzed using liquid chromatography coupled with tandem mass spectrometry (LC-MS). Data were acquired with data-independent acquisition parallel accumulation– serial fragmentation (diaPASEF) and processed using DIA-NN software. Differentially expressed proteins informed model development. Random Forest models were trained in a Discovery cohort (n=55) to select clinical variables linked to the proteome, then tested in an independent Validation cohort (n=59). Recursive feature elimination (RFE) identified a minimal feature set that was predictive of sepsis. The performance was assessed using repeated cross-validation and external validation.

Results

Twelve plasma proteins differed between sepsis and non-sepsis patients at FDR < 0.1, corresponding to 26 proteome-enriched clinical variables. The classifier achieved mean AUC’s of 0.73 and 0.76 in Discovery and Validation cohorts, respectively. RFE performance plateaued with ≥9 variables, peaked at an accuracy of 0.78, and deteriorated below seven; the final three features before collapse were plasma BUN, chemokine ligand 3 (CCL3), and creatinine. Proteome-to-clinical regression highlighted creatinine as having the strongest correlation (R² = 0.558). Discussion : A concise set of routinely obtainable variables anchored by renal markers and CCL3 captured proteomic signals and discriminated sepsis across cohorts, supporting a “proteomics-informed, clinic-first” strategy for pragmatic EHR deployment.

While larger multicenter studies are warranted, these findings suggest that renal dysfunction exerts a disproportionate influence on sepsis and that increased emphasis on kidney-related markers may improve both recognition and risk assessment.