OBJECTIVES
The aim of this study was to assess the performance of a multimarker model in distinguishing patients with sepsis from those with non-infective systemic inflammatory response.
METHODS
This study is part of a prospective study of community-onset severe sepsis and septic shock in adults conducted from September 2011 to June 2012 at Skaraborg Hospital, in the western region of Sweden. The levels of 92 inflammation-related human protein biomarkers were measured simultaneously using Proseek® Multiplex Inflammation I96x96 (Olink Bioscience, Sweden) in 122 plasma samples collected from patients suspected with sepsis. After pre-processing normalization procedure, measurements of the markers were obtained as Normalized Protein eXpression (NPX) units on a log2 scale (GenEx, MultiD Analyses AB, Sweden). The study was approved by the Regional Ethical Review Board of Gothenburg (376-11). All patients enrolled provided written informed consent.
To reduce the number of markers, factor analysis was performed. Thereafter, a multimarker model for classification was derived using discriminant analysis. The multimarker model consisted of a linear function of the selected markers. Cross-validation was performed by classifying each sample by the discriminant function derived from all samples other than that specific sample. The performance was assessed as area under receiving operating characteristic (ROC) curve. The cut-off for sensitivity and specificity was derived from the cut score of the discriminant function. Statistical analyses were performed in SPSS 22.0 (IBM Corporation Somers, NY USA).
RESULTS
Of the 122 samples, 80 (66%) were from patients diagnosed with sepsis and 42 from patients with non-infective systemic inflammatory response syndrome (SIRS). The five markers selected for the multimarker model were interleukin-6 (IL-6), cystatin D (CST5), delta and notch-like epidermal growth factor-related receptor (DNER), STAM-binding protein (STAMPB), macrophage colony-stimulating factor 1 (CSF 1). Every single marker was statistically different between the groups (p value < 0.001), except for DNER (p value 0.064) and STAMPB (p value 0.060). The area under ROC was higher for the multimarker model (81%) than for each biomarker separately (Figure 1). The accuracy for the multimarker model was 72% [64-80, 95% CI]; sensitivity 84% [77-91, 95% CI]; specificity 60% [51-69, 95% CI]; positive predictive value 79% [72-86, 95% CI]; and negative predictive value 66% [58-74, 95% CI].
CONCLUSION
A higher power of discrimination is obtained by combining more than one biomarker. However, the multimarker candidates identified in this study need further assessment.