Principles for high-quality, high-value testing

Abstract

A survey of doctors working in two large NHS hospitals identified over 120 laboratory tests, imaging investigations and investigational procedures that they considered not to be overused. A common suggestion in this survey was that more training was required. And, this prompted the development of a list of core principles for high-quality, high-value testing. The list can be used as a framework for training and as a reference source. The core principles are: (1) Base testing practices on the best available evidence. (2) Apply the evidence on test performance with careful judgement. (3) Test efficiently. (4) Consider the value (and affordability) of a test before requesting it. (5) Be aware of the downsides and drivers of overdiagnosis. (6) Confront uncertainties. (7) Be patient-centred in your approach. (8) Consider ethical issues. (9) Be aware of normal cognitive limitations and biases when testing. (10) Follow the 'knowledge journey' when teaching and learning these core principles.

Figures

Figure 1

Reclassification possibilities with the last in a series of tests. This figure is only reproduced in colour in the online version.

Figure 2

Receiver operating characteristic (ROC) curves for C reactive protein (CRP) and procalcitonin (PCT) levels for serious infection. The area between the ROC curves represents the net reclassification index (NRI), which is negative for PCT compared to CRP. The graph is from van den Bruel et al, and reproduced with permission from the publishers. This figure is only reproduced in colour in the online version.

Figure 3

Rules of thumb for testing when sensitivity and specificity are 80–90%, and positive and negative likelihood ratios 4–9 and 0.3–0.1. The horizontal line shows the threshold for action. Upward-sloping lines point to positive predictive values. Downward-sloping lines point to negative predictive values. The angles of the prediction lines reflect the likelihood ratios. Thick prediction lines show results that change management. Thin prediction lines show results that will not change management. The moderate slopes of the prediction lines reflect the combination of moderately high sensitivity and moderately high specificity. Prevalence categories are labelled ‘Don't test’ if the result of testing will not change management. This figure is only reproduced in colour in the online version.

Figure 4

SpIn—rule of thumb for using a test with high specificity and low sensitivity. For example, HbA1c≥6.5% for diagnosing diabetes has 99% specificity and 30% sensitivity, and positive and negative likelihood ratios 30 and 0.7. The horizontal line shows the threshold for action. Upward-sloping lines point to positive predictive values. Downward-sloping lines point to negative predictive values. The angles of the prediction lines reflect the likelihood ratios. Thick prediction lines show results that change management. Thin prediction lines show results that will not change management. The gentle downward and steep upward slopes of the prediction lines reflect the combination of low sensitivity and high specificity. This figure is only reproduced in colour in the online version.

Figure 5

SnOut—rule of thumb for using a test with high sensitivity and low specificity. For example genetic typing for coeliac disease has 99% sensitivity and 54% specificity, and positive and negative likelihood ratios 2.2 and 0.02. The horizontal line shows the threshold for action. Upward-sloping lines point to positive predictive values. Downward-sloping lines point to negative predictive values. The angles of the prediction lines reflect the likelihood ratios. Thick prediction lines show results that change management. Thin prediction lines show results that will not change management. The steep downward and gentle upward slopes of the prediction lines reflect the combination of high sensitivity and low specificity. This figure is only reproduced in colour in the online version.

Figure 6

The blue line shows how the post-test probability of HIV rises as more diagnostic information becomes available from the history, examination and laboratory tests. In contrast to figures 2–4, the vertical scale is logarithmic in order to expand the scope of the extremes of the axis. With log scaling, the graph can show how the final test (HIV western blot) limits the false-positive rate to less than around 1 in 10 000—anything less than this is unacceptable for diagnosing HIV. The Figure is adapted from Henríquez. Probabilities in the labels are qualitatively, but not quantitatively, accurate. This figure is only reproduced in colour in the online version.