March 24, 2017

Present exaggerated, future hypothetical

There’s a story at Stuff with the headline “Blood test picks up cancer before symptoms start to show“. It’s attributed to the Telegraph but doesn’t seem to have run there yet.( It’s here. I don’t think it was up when I wrote that, but I might have missed it).  The test looks at tiny amounts of tumour DNA that are circulating in the blood.

It’s not straightforward to demonstrate that a new test is better for early cancer detection than current methods. You need to use the test on a group of people who have cancer but don’t know it yet. Recruiting them is hard. In this story, the researchers didn’t do that. They made imaginary people with undetected cancer in a computer, and tested their imaginary blood to see if the cancer would be picked up.  That’s a good idea, but it’s a very early stage in the research process.

The researchers also went on to use the method on some real blood samples of people with lung, liver, or breast cancer, and without cancer. These aren’t the currently-undiagnosable people who might benefit in the future, but most of the cancer patients had early-stage disease by current standards.

The story says these results were very good

It picked up eight out of 10 cancers, and gave a false positive on fewer than one in 100 occasions.

The press release and the paper itself aren’t as positive. The test did pick up 8 out of 10 early-stage cancers, but the false positive rate was 20%.  That’s far too high for a screening test in people with no symptoms.  The test also did a lot better on liver cancer than on lung cancer or breast cancer, which isn’t very helpful for NZ or for the UK, since liver cancer is relatively rare in both countries.

As the story does say, this research is in its infancy, and the new test did much better than some standard machine-learning methods run on the same data. But it’s nowhere near ready for prime time: the headline is referring to potential future benefits of a different, further-developed test, but is using the present tense.

 

avatar

Thomas Lumley (@tslumley) is Professor of Biostatistics at the University of Auckland. His research interests include semiparametric models, survey sampling, statistical computing, foundations of statistics, and whatever methodological problems his medical collaborators come up with. He also blogs at Biased and Inefficient See all posts by Thomas Lumley »