November 21, 2018 | 0 Feedbacks on “ROC Curve Analysis” | Biostatistics & Bioinformatics | Brianne
ROC Curve Analysis
What is it? A Receiver's Operating Characteristic (ROC) curve plots every value of a continuous measurement by its specificity and sensitivity to distinguish health status in a population of subjects. The area under the curve (AUC) reflects the measurement's potential to be a diagnosis tool. At a specific sensitivity, the specificity can be determined, or vice versa.
When is it used? This analysis is used to identify the appropriate classifying thresholds to diagnose a patient with expected sensitivity and specificity.
How does it work?
ROC curve analysis: Example
We've identified a potential biomarker, Protein "A", of Alzheimer's disease that is elevated in the plasma of Alzheimer's patients compared to healthy patients (Figure 1). We now need to determine the lower cut-off value of "Protein A" levels that will identify a patient with Alzheimer's disease. We don't want to have the threshold too low like concentration X in Figure 1 or else a lot of healthy patients will be wrongly diagnosed. However we also don't want to have a threshold that is really high like concentration Y in Figure 1 because a lot of Alzheimer's patients won't get diagnosed. A ROC curve can help identify the "sweet spot" (i.e., optimum sensitivity-specificity balance).
Figure 2 explains what sensitivity and specificity are. Ideally, the sensitivity and specificity would be 100%. In reality, virtually all biomarkers do not have perfect sensitivity and specificity.
A ROC curve is generated across all values and the AUC is determined (Figure 3). Higher AUC values represent a better biomarker. A point along the ROC curve is chosen with the desired trade-off between sensitivity and specificity. With known sensitivity and specificity, the cut-off value can be ascertained.
For this example, let's assume the black dashed line is the ROC curve for our data. We would likely choose the X-Y coordinate of 0.1, 0.8, such that the biomarker would have a specificity of 90% and a sensitivity of 80%.



What does the data look like? ROC curve analyses are usually portrayed as a plot like Figure 3.