360 likes | 772 Views
ROC. Receiver Operating Characteristic- historic name from radar studies Relative Operating Characteristic - psychology, psychophysics Operating Characteristic - preferred by some. Disease -. Disease +. TN. TP. FN. FP. Test -. Test +. Specificity TNF = TN/(TN+FP) 0.86
E N D
ROC Receiver Operating Characteristic- historic name from radar studies Relative Operating Characteristic - psychology, psychophysics Operating Characteristic - preferred by some
Disease - Disease + TN TP FN FP Test - Test + Specificity TNF = TN/(TN+FP) 0.86 TNF + FPF = 1 Sensitivity TPF = TP/(TP+FN) 0.73 TPF + FNF = 1 Threshold for Positivity
Test Characterization • SENSITIVITY of a test is its ability to detect disease within a diseased population. It is calculated as the fraction of diseased patients correctly identified by the test. Also called the True Positive Fraction and TPF= TP/(TP+FN) where (TP + FN) is the number of patients with the disease. • SPECIFICITY of a test is its ability to identify the absence of disease in a disease free population. It is calculated as the fraction of non-diseased patients correctly identified by the test. Also called True Negative Fraction and TNF= TN/(TN + FP) where (TN + FP) is the number of patients that are disease free. • ACCURACY is the fraction of correct test results or diagnoses. It is calculated as the number of patients with correct test results divided by the whole patient population (TP+TN)/(TP+FP+TN+FN). • PREVALENCE of the disease is is calculated as the fraction of patients who have the disease (TP+FN)/(TP+FP+TN+FN).
Example: A study shows 90 true positives, 80 false positives, 20 true negatives and 10 false negatives. What are the sensitivity and specificity of the test? Sensitivity = TPF TPF = TP/(TP + FN) TPF = 90/(90 +10) = 0.90 Sensitivity = 90% Specificity = TNF TPF = TN/(TN + FP) TPF = 20/(20 +80) = 0.20 Specificity = 20% Accuracy = ?
ROC Curves • The receiver operating characteristic (ROC) curve is used to compare overall performance (sensitivity and specificity) of a test. It can be used where imaging systems and observers (radiologists) both play a role in the test outcome or with purely objective tests such as with blood levels of cholesterol (LDL and or HDL). • An ROC curve is a graph of the True Positive Fraction (sensitivity) vs. False Positive Fraction (1-specificity) of a test as the threshold for positive result is changed.
ROC Curve TPF FPF D- mean = 150, SD = 50 D+ mean = 250, SD = 75
ROC - Diagnostic Imaging Threshold criteria are established using an ordinal scale of 0-4, ranging from under-reading (0) to over-reading (4). • At the most restrictive criterion (under reading or high threshold for positive), both sensitivity and the false-positive fraction are near zero (lower left on ROC). • At the most lax criterion (over reading or low threshold for positive), both the sensitivity and the false-positive fraction are near 1 (upper right on ROC). • In practice the operating point is a compromise between increasing sensitivity and decreasing specificity (increasing FPF).
Hypothetical ROC curve 1.0 An experienced radiologist Over-reading Less experiencedreader True positive fraction (sensitivity) Useless Test How can this be? Under-reading 0 1.0 False positive fraction (1 - specificity)
Non-diseased cases TPF, sensitivity Threshold less aggressive mindset Diseased cases FPF, 1-specificity
Non-diseased cases moderate mindset TPF, sensitivity Threshold Diseased cases FPF, 1-specificity
Non-diseased cases more aggressive mindset TPF, sensitivity Threshold Diseased cases FPF, 1-specificity
Threshold Non-diseased cases Entire ROC curve TPF, sensitivity Diseased cases FPF, 1-specificity
Entire ROC curve chance line TPF, sensitivity Reader Skill and/or Level of Technology FPF, 1-specificity
Quantifying ROC Curves • The area under an ROC curve is a measure of overall performance. • The maximum area is 1.0 • Useless test is the diagonal line from 0.0 to 1.0 and has area under ROC =0.5, so a more meaningful measure is the area in excess of 0.5. • As test performance improves, the curve moves towards the upper left corner and the area under ROC increase.
Dilemma:Which modality is better? 1.0 Modality B Modality A True Positive Fraction 0.0 0.0 1.0 False Positive Fraction
ROCs (one outcome) 1.0 B is better than A Modality B True Positive Fraction Modality A 0.0 0.0 1.0 False Positive Fraction
ROC (another outcome) 1.0 B same as A Modality B Modality A True Positive Fraction 0.0 0.0 1.0 False Positive Fraction
ROC (another outcome) 1.0 A is better than B Modality B Modality A True Positive Fraction 0.0 0.0 1.0 False Positive Fraction