ROC analysis has been mainly used to assess diagnostic checks.
The ROC analysis starts with the observer assigning exactly two classifications to a case after observing one stimulus associated with the case. The two classification states are whether there is presence of a particular disease or not and normal and abnormal. Bases on the information from imaging exam the observers task is to determine the correct disease classification. Then the observers classification is compared to the true classification of the reference variable. Subsequently, sensitivity and specificity measures are constructed. Sensitivity is the proportion of correctly specified cases among those who are truly positive. Specificity is the proportion of correctly specified cases among those who are truly negative. In real world classification is associated with uncertainty therefore sensitivity and specificity can never be 100. The trade-off between sensitivity and specificity depends on the threshold set by the observer. Low threshold shows tendency to over-call high sensitivity relatively to low specificity. On the other hand higher threshold shows the tendency to under-call which refers to low sensitivity to relatively low specificity.The horizontally flipped plot of sensitivity over specificity constitutes the so called ROC curves graph. The area under the ROC curve (AUC) is the global indicator of the diagnostic performance of the test and represents the probability of the observer correctly identifying the positive case with a randomly chosen pair of cases where one is positive and the other one is negative. AUC also represents the average over entire range of possible sensitivities. The most important property of ROC curves is independence from the threshold set by the observer when choosing. It is a more accurate measure as it covers all possible interpretation thresholds. Parametric methods are used when the statistical distribution of test values in diseased and non-diseased is known. Binormal distribution is commonly used for this purpose. This is applicable when both diseased and non-diseased test values follow normal distribution. If data are actually binormal or a transformation such as log, square or Box- Cox makes the data binormal then the relevant parameters can be easily estimated by the means and variances of test values in diseased and non-diseased subjects. Figure 1
Evaluation of optimal BMI cut-off point.
Three criteria are used to find optimal threshold point from ROC curve. First two methods give equal weight to sensitivity and specificity and impose no ethical, cost, and no prevalence constraints. The third criterion considers cost which mainly includes financial cost for correct and false diagnosis, cost of discomfort to person caused by treatment, and cost of further investigation when needed. This method is rarely used in medical literature because it is difficult to estimate the respective costs and prevalence is often difficult to assess. These three criteria are known as the points on curve closest to the (0, 1)(shortest distance (C) , Youden index (J), and minimize cost criterion, respectively. Based on model 1, the first criterion the optimal cut-off points of the classification variable in the ROC analysis can be calculated from the shortest distance between the ROC curve and the left hand side of the ROC curve where sensitivity and specificity equal 1. If sn and sp denote sensitivity and specificity, respectively, the distance between the point (0, 1) and any point on the ROC curve is d = [(1 sensitivity) 2+ (1 specificity)2.To obtain the SN= Sensitivity (True positive rate), SP Specificity (False positive rate) optimal cut-off point to discriminate the disease with non-disease subject, we need to calculate this distance for each observed cut-off point, and locate the point where the distance is minimum. The second is Youden index that maximizes the vertical distance from line of equality to the point B[x, y] as shown in the Figure 1 . The x-axis represents (1specificity) and y-axis represents sensitivity. In other words, the Youden index J is the point on the ROC curve which is farthest from line of equality (diagonal line). The main aim of Youden index is to maximize the difference between TPR (sensitivity) and FPR (1 specificity) and little algebra yields J = max [sensitivity (B)+specificity(B)-1]. The value of J for continuous test can be located by doing a search of plausible values where sum of sensitivity and specificity can be maximum. Youden index is more commonly used criterion because this index reflects the intension to maximize the correct classification rate and is easy to calculate. The index is defined for all points of an ROC curve, and the maximum
value of the index may be used as a criterion for selecting the optimum cut-off point when a diagnostic test gives a numeric rather than a dichotomous result.