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 (R squared of 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.
Ardabili et al. (Wed,) studied this question.