Proteomics improves the prediction of burns mortality: results from regression spline modeling

Abstract

Prediction of mortality in severely burned patients remains unreliable. Although clinical covariates and plasma protein abundance have been used with varying degrees of success, the triad of burn size, inhalation injury, and age remains the most reliable predictor. We investigated the effect of combining proteomics variables with these three clinical covariates on prediction of mortality in burned children. Serum samples were collected from 330 burned children (burns covering >25% of the total body surface area) between admission and the time of the first operation for clinical chemistry analyses and proteomic assays of cytokines. Principal component analysis revealed that serum protein abundance and the clinical covariates each provided independent information regarding patient survival. To determine whether combining proteomics with clinical variables improves prediction of patient mortality, we used multivariate adaptive regression splines, because the relationships between analytes and mortality were not linear. Combining these factors increased overall outcome prediction accuracy from 52% to 81% and area under the receiver operating characteristic curve from 0.82 to 0.95. Thus, the predictive accuracy of burns mortality is substantially improved by combining protein abundance information with clinical covariates in a multivariate adaptive regression splines classifier, a model currently being validated in a prospective study.

Figure 1

Principal component analysis. Loading of each variable on component 2 is plotted on the vertical dimension, whereas loading on component 1 is plotted on the horizontal dimension. Note that the two dimensions pass through two major clusters: proteomics variables and clinical variables [total burn surface area (TBSA), percentage third‐degree burn (percentage 3D), inhalation injury (INJ), and age].

Figure 2

Receiver operating characteristic (ROC) analysis. ROC curves are shown for MARS modeling using clinical features (filled circle) and a combination of proteomics data and clinical features (open circles). Note that the ROC curve for the combined MARS model is shifted upward and toward the upper left quadrant.