Fig 1.
The workflow of the study began with the collection of blood samples and ended with the identification and evaluation of 6 metabolite variables for prediction of bacteremic sepsis.
Table 1.
Characteristics and clinical variables of patients.
Fig 2.
Multivariate data analysis with work set samples (n = 72) using 107 metabolites.
Panel A shows a PCA plot of the 4/5th principal component separating bacteremic sepsis (black circles) from ER controls (white circles). Panel B shows an OPLS-DA plot of t1/t(o)1, discriminating bacteremic sepsis (black circles) from ER controls (white circles), (P = 4.8×10−12).
Fig 3.
Scatter plot of class discrimination based on cross-validated scores in an OPLS-DA model using 107 metabolites.
The samples of the work set with bacteremic sepsis (black circles) and the work set of ER controls (open circles) were used for modeling. Validation using the samples of the test set with bacteremic sepsis (black triangles) and the test set of ER control (open triangles) is shown. The Y-axis represents the seven fold cross-validated predictive score vector 1. Error bars represent mean score values with 95% confidence intervals.
Fig 4.
OPLS-DA models using subsets of metabolites and predictions of test set samples.
Panel A shows an OPLS-DA plot using work set samples and 38 metabolites discriminating bacteremic sepsis (black circles) from ER controls (open circles), (P = 2.0×10−17). Panel B shows a scatter plot of the class discrimination using cross-validated scores (tPS[1]cv[7]p) of an OPLS-DA model with 38 metabolites. Work set bacteremic sepsis (black circles), work set ER controls (open circles), test set bacteremic sepsis (black triangles) and tests set ER control (open triangles) are shown. Panel C shows an OPLS-DA plot using work set samples and 6 metabolites discriminating bacteremic sepsis (black circles) from ER controls (open circles), (P = 4.1×10−11). Panel D shows a scatter plot of the class discrimination using cross-validated scores (tPS[1]cv[7]p) of an OPLS-DA model with 6 metabolites. Work set bacteremic sepsis (black circles), work set ER controls (open circles), test set bacteremic sepsis (black triangles) and tests set ER control (open triangles) are shown. Error bars in panel B and D represent mean score values with 95% confidence intervals.
Fig 5.
The regression coefficient plot for the OPLS-DA model with 38 metabolites using work set samples.
Positive regression coefficients indicate a positive correlation with bacteremic sepsis and negative coefficient a negative correlation.
Fig 6.
ROC curves of metabolites and laboratory diagnostic variables available in the clinic for the prediction of bacteremic sepsis.
Panel A shows logistic regression modelling on work set samples using 6 metabolites (solid line), the combination of temperature, C-reactive protein, thrombocyte, and white blood cell count (dotted line) and the combination score SIRS ≥2 (dashed line). Panel B shows prediction on test set samples using 6 metabolites (solid line), the 4 best clinical variables (dotted line) and the combination variable SIRS ≥2 (dashed line). Panel C shows logistic regression modelling on single variables of work set samples for myristic acid (solid line), white blood cell count (dotted line) and C-reactive protein (dashed line). Panel D shows prediction on test set samples using myristic acid (solid line), white blood cell count (dotted line) and C-reactive protein (dashed line).
Table 2.
AUC and model performance for work set and test set.
Fig 7.
Tukey’s box-and-whisker plots for work set samples using the 6 most important metabolites.
Values are corresponding to chromatogram peak areas. Outliers are represented by dots outside the 1.5 interquartile range of the 25 respective 75 percentile. Mean values are indicated by a plus sign. The P values were derived from MannWhitneyU tests *P < .05, **P < .01, ***P < .005, ****P < .001.