medical and biological sciences - Collegium Medicum - Uniwersytet ...

Measures of diagnostic accuracy: basic definitions 63with a different prevalence of the disease in the population.Nonetheless, sensitivity and specificity can varygreatly depending on the spectrum of the disease in thestudy group.PREDICTIVE VALUESPositive predictive value (PPV) defines the probabilityof having the state/disease of interest in a subjectwith positive result (B+|T+). Therefore PPV representsa proportion of patients with positive test result intotal of subjects with positive result (TP/TP+FP).Negative predictive value (NPV) describes theprobability of not having a disease in a subject with anegative test result (B-|T-). NPV is defined as a proportionof subjects without the disease with a negative testresult in total of subjects with negative test results(TN/TN+FN).Unlike sensitivity and specificity, predictive valuesare largely dependent on disease prevalence in examinedpopulation. Therefore, predictive values from onestudy should not be transferred to some other settingwith a different prevalence of the disease in the population.Prevalence affects PPV and NPV differently. PPVincreases, while NPV decreases with the increase ofthe prevalence of the disease in a population. Whereasthe change in PPV is rather substantial, NPV is somewhatless influenced by the disease prevalence.LIKELIHOOD RATIO (LR)Likelihood ratio is a very useful measure of diagnosticaccuracy. It is defined as the ratio of expectedtest result in subjects with a certain state/disease to thesubjects without the disease. As such, LR directly linksthe pre-test and post-test probability of a disease in aspecific patient [3]. Simplified, LR tells us how manytimes more likely particular test result is in subjectswith the disease than in those without disease. Whenboth probabilities are equal, such test is of no value andits LR = 1.Likelihood ratio for positive test results (LR+) tellsus how much more likely the positive test result is tooccur in subjects with the disease compared to thosewithout the disease (LR+=(T+│B+)/(T+│B-)). LR+ isusually higher than 1 because it is more likely that thepositive test result will occur in subjects with the diseasethan in subject without the disease. LR+ can besimply calculated according to the following formula:LR+ = sensitivity / (1-specificity).LR+ is the best indicator for ruling-in a diagnosis.The higher the LR+ the more indicative of a diseasethe test is. Good diagnostic tests have LR+ > 10 andtheir positive result has a significant contribution to thediagnosis.Likelihood ratio for negative test result (LR-)represents the ratio of the probability that a negativeresult will occur in subjects with the disease to theprobability that the same result will occur in subjectswithout the disease. Therefore, LR- tells us how muchless likely the negative test result is to occur in a patientthan in a subject without disease. (LR-=(T-│B+)/(T-│B-)). LR- is usually less than 1 because it isless likely that negative test result occurs in subjectswith than in subjects without disease. LR- is calculatedaccording to the following formula: LR- = (1-sensitivity) / specificity.LR- is a good indicator for ruling-out a diagnosis.Good diagnostic tests have LR- < 0,1. The lower theLR- the more significant contribution of the test is inruling-out, i.e. in lowering the posterior probability ofthe subject having the disease.Since both specificity and sensitivity are used tocalculate the likelihood ratio, it is clear that neitherLR+ nor LR- depend on the disease prevalence inexamined groups. Consequently, the likelihood ratiosfrom one study are applicable to some other clinicalsetting, as long as the definition of the disease is notchanged. If the way of defining the disease varies, noneof the calculated measures will apply in some otherclinical context.ROC CURVEThere is a pair of diagnostic sensitivity and specificityvalues for every individual cut-off. To constructa ROC graph, we plot these pairs of values on thegraph with the 1-specificity on the x-axis and sensitivityon the y-axis (Figure 1).Fig. 1. ROC curve