Masters candidate in Biostatistics, Yoon Son Ahn, will present:
“Optimizing Indices for a Disease Biomarker”
Plan B Adviser: Chap Le
Abstract: Even though an ROC curve is a useful plot to show the validity of a test, it does not suggest an optimal cutpoint to outline the range of biomarker values with disease. To investigate this aim, we explored response difference (Youden’s index), distance to best corner of ROC curve, distance to worst corner of ROC curve, diagnostic odds ratio, and relative response to see if these five indices suggest the same optimal cutpoint with a prostate cancer dataset. The dataset has 101 patients that 50 are healthy and 51 have prostate cancer. The following are the results of the indices: Youden’s index (J) has its optimal cutpoint at 10.6 (sensitivity=0.373, specificity=0.920). The distance to best corner of ROC curve (d) is minimized when PSA is split at 5.75 (sensitivity=0.686, specificity=0.520). On the other hand, the distance to worst corner of ROC curve (D) is maximized at 17.85 (sensitivity=0.196, specificity=1.000). Relative response (RR) has an optimal cutpoint at 16.5 (sensitivity=0.196, specificity=0.980). Lastly, odds ratio (OR) is maximized at 16.5 (sensitivity=0.196, specificity=0.980). Relative response and odds ratio have the same optimal cutpoint, but other indices have different optimal cutpoints. Further implications, we would suggest three approaches: finding the smallest standard error among the indices, building a regression model with adjustments of subject characteristics, or building a model for the ROC curve.
Keywords: disease biomarker cutpoint, optimal cutpoint, optimization from ROC curve, sensitivity, and specificity.
Refreshments will be served prior to the presentation.