Spectrum bias: a quantitative and graphical analysis of the variability of medical diagnostic test performance.

摘要: The performance of a medical diagnostic test may vary in subgroup patients according to the severity and clinical presentation disease. This phenomenon is often referred as spectrum effect. There currently no technique determine which situations this effect lead bias, that is, distortion posterior probability can potentially affect decision. We show on either positive or negative result, be expressed subgroup-specific likelihood ratio (LR) divided by LR overall population patients. assessment bias independent pretest probability. present here statistic its variance also propose mode graphically visualizing presence bias. applied it 15 examples from literature, three are discussed detail. In two examples, there was bias: (1) diagnosis urinary tract infection subgroups low high risk when using dipstick test, (2) left ventricular hypertrophy electrocardiography subjects with different body mass index. third example, for coronary heart disease defined age, sex comorbidities exercise electrocardiography. As sensitivity specificity usually opposite directions one other, LRs tends remain constant, resulting little conclude more common than decisions affected only under certain conditions.