Proteomics in graft-versushost disease and graftversus-leukemia - Immune Biology of Allogeneic Hematopoietic Stem Cell Transplantation

Biology Reference

In-Depth Information

458

FIGURE 19.3

Receiver operating characteristics (ROC) curves for diagnosis of acute GVHD. An ROC curve is a plot of the true-positive rate on the y axis (sensitivity) versus the

false-positive rate (1 − specificity) on the x axis of a given biomarker. (A) A perfect test would result in an ROC curve appearing as a right angle, indicating that 100%

of the samples are true as opposed to false positives. (B) In the perfect case, the corresponding area under the curve (AUC) will equal 1. (C) A random test will have an

AUC of 0.5, meaning that there is one false-positive for every true-positive sample. (D) Individual ROC curves for IL-2R α , TNFR1, HGF, and IL-8 and the composite panel

comparing GVHD− ( N = 166) vs GVHD+ ( N = 116), with a threshold determined by a linear combination of values for IL-2R α , TNFR1, HGF, and IL-8 concentrations in the

training set. The AUC is 0.91. (E) ROC curve for elafin comparing patients post-HSCT with skin GVHD ( N = 159) vs patients post-HSCT with rashes of other etiology ( N =

53); the threshold used is the median of patients with skin GVHD, the AUC is 0.73. (F) ROC curve of patients with lower GI GVHD ( N = 162) vs patients post-HSCT with

diarrhea of other etiology ( N = 42); the threshold is the median of patients with lower GI GVHD. The AUC is 0.80.

predictive, diagnostic performance, as was the case in our first biomarker

panel [12] . To create a comprehensive GVHD biomarker panel, we used

a proportional odds logistic regression model to determine a compos-

ite panel that will generate an ROC curve with an area under the curve

(AUC) >0.8, meaning that 80% of the positive results are true positives.

Presumably, for an aGVHD diagnosis, a combination of tissue-specific and

systemic biomarkers will be more informative than individual markers.

However, if a biomarker is not highly correlated with either other biomark-

ers or clinical predictors, one or two biomarkers could be sufficient for

either diagnostic or predictive tests. To best evaluate the number of bio-

markers that will give the most information, we used optimized classifica-