Biases that make screening programs look good

Too often lung cancers are diagnosed at a late, untreatable stage, as was the cancer in the CT above. An effective lung cancer screening tool would be great!


Last week I described a recent trial that disappointed, failing to prove that lung cancer screening with low-dose CT (LDCT) saves lives. The designers of this trial are to be applauded for their trial design. They have understood, and tried to avoid, three forms of bias which can make a screening program looks like it is working - when it isn't. These biases have had an impact in the field of prostate and breast cancer screening also - where recent evidence suggests that benefits of screening in some populations (for example women aged 40-50) may have been overstated.

Imagine two people with very small, slow growing, early lung cancers attend different hospitals with indigestion, and are seen by different doctors who manage them independently. One doctor, on the barest provocation, arranges a CT chest for his patient and a small tumour is seen. The other patient does not undergo CT chest. After a run of investigations the first patient has his cancer removed and is followed up for 5 years before being declared free of disease. The second patient is discharged home, but 5 years later presents to his doctor coughing up blood - and a large lung cancer is found on chest xray. He too is treated surgically. Two years later, both patients present to their independent doctors with weight loss and abdominal pain. They are found to have distant spread of cancer and both die 6 months later. The first patient, it seems, survived 7.5 years from the time of diagnosis, the second only 2.5 years. Increasing survival from diagnosis for a lung cancer looks like a good thing, but in this instance survival may not really have been increased; both had cancer at time zero, but only one was diagnosed. This lead to an appearance of increased survival when the cancer was just diagnosed earlier. We call this lead time bias.

The second form of bias is like it. Suppose that a screening program only picks up slow-growing tumours. Suppose aggressive tumours generally make their presence felt clinically within about 6 months and are picked up clinically, and therefore not on screening CT. If a trial predominantly picks up slow-growing tumours then it is possible that a trial will seem to indicate increased survival time in patients undergoing screening only because they are the ones with all the slow-growing tumours. this is called 'length bias'.


Finally, lung cancer prevalence increases with age. It is likely that a number of older people die with lung cancer, rather than from it. This was certainly the finding in the recent Dante trial, the subject of the previous blog. A trial of LDCT screening may pick up, and 'cure', a number of indolent tumours that were never going to make their presence felt anyway. This is 'overdiagnosis bias'.

To overcome these biases it is very important that trials of screening programs are run in a randomised and controlled way. There must be a control group managed in 'usual' fashion. The results of a screening program cannot be compared to historical controls/cancer survival data. And they must be run over an adequate number of years to demonstrate if there really is an overall survival benefit. After all, a screening program that doesn't result in less people really dying from lung cancer, while it may serve to employ lots of people in medical jobs, is in essence a waste of time.

Andrew