Evaluation of the accuracy of medical tests in a region around the optimal point.

RATIONALE AND

OBJECTIVES

The accuracy of medical tests is often assessed using the area under the entire receiver-operating characteristic (ROC) curve. However, this includes values that might be of no clinical importance. Evaluation of a portion of the curve, or a single point, requires identifying a range of clinical interest, which may not be obvious. The author suggests evaluating the accuracy of medical tests in the vicinity of the optimal point.

MATERIALS AND

METHODS

Assuming binormality, the author estimated the optimal threshold as the value that maximizes the generalized Youden index. The confidence interval around the optimal point defined a region of clinical interest; the accuracy of the medical test was assessed using the partial area index (PAI) and standardized partial area (sPA). Bootstrapping was used to estimate variances and construct confidence intervals. Coverage probabilities for the PAI and sPA were assessed, as was the size of the test to compare measures. An example using biomechanical measures from radiographic images of the pelvis and lumbar spine to detect disk hernia and spondylolisthesis is presented.

RESULTS

Coverage probabilities of confidence intervals for the partial area measures were good. The size of the test to compare partial area measures was appropriate. Values of PAI and sPA varied with the cost/prevalence ratio. In the example, the biomechanical measures were not found to have significantly different accuracy around the optimal point.

CONCLUSIONS

The PAI and sPA associated with the optimal point were found to be reasonable and useful measures of accuracy.