Redundancy of single diagnostic test evaluation

Diagnostic research and diagnostic practice frequently do not cohere. Studies commonly evaluate whether a single test discriminates between disease presence and absence, whereas in practice a test is always judged in the context of other information. This study illustrates drawbacks of single-test evaluation and discusses principles of diagnostic research. We used data on 140 patients suspected of pulmonary embolism who had an inconclusive ventilation-perfusion lung scan. We evaluated three tests: partial pressure of oxygen in arterial blood (PaO2), x-ray film of the thorax, and leg ultrasound. On the basis of single-test evaluations, ultrasound was most informative. Given a prior probability of 0.27, it had a much better combination of positive and negative predictive value (0.71 and 0.21, respectively) relative to thorax x-ray (0.33 and 0.11) and PaO2 (0.35 and 0.27). The combination of positive and negative likelihood ratio was also more promising for ultrasound (7.3 and 0.7) than for thorax x-ray (1.3 and 0.3) and PaO2 (1.3 and 0.9). As the tests are always performed after the history and physical, we judged their added value using multivariable logistic modeling with receiver operating characteristic (ROC) analyses. The ROC areas of the model, including history and physical, with additional PaO2, thorax x-ray, or ultrasound, were 0.75, 0.77, 0.81, and 0.81, respectively, which indicated similar added value of thorax x-ray and ultrasound. Application of the models to patient subgroups also yielded added predictive value for thorax x-ray film. Thus, the results of single-test evaluations may be very misleading. As no diagnosis is based on one test, single-test evaluations have limited value in diagnostic research and only have relevance in the context of screening and the initial phase of test development. Diagnostic research should always apply an approach of constructing, extending, and validating diagnostic models in agreement with routine clinical work-up using logistic regression analyses.