Back in the dawn of home computing, there was a war between those who were fans of the colourful version of the ZX Spectrum and those who were loyal to the Vic20s new big brother, the C64. The problem of spectrum bias is not in any way related to these forerunners of today's games consoles. It is, though, a potential problem when we evaluate diagnostic tests across different healthcare settings.When researchers are assessing how effective diagnostic tests are, they will compare the ability to make or exclude a diagnosis with a reference standard in a population similar to which they hope it will be used in practice. This group of those with positive tests will be a mix of people with the target condition and those without (a false positive; show positive on the test without the disease being present). Of course there will be those with negative tests—those who have the disease (false negatives) and those without the disease (true negatives) ... Unless the new test is perfect in every way, but Mary Poppins Tests are fantastical.The reason why we get false positives, and false negatives, might vary across the different groups of patients.For example, it might be that white cells in the urine are a great indicator of potential urinary tract infection, but if you use this test in a neutropenic population, with a very low white count, it loses its ability. Similarly, if you use the physical examination finding of wheeze in a group of kids at high risk of invasive fungal infection you might find a load of false positives. This is the issue of spectrum bias, and while an ideal way around would be to evaluate every diagnostic test in every possible population, a more pragmatic approach is to use clinical expertise and smidgens of data to understand how test characteristics may be blunted or sharpened in different areas of practice.
展开▼
