New lung cancer prediction tool promises better use of screening CT

New Prediction Model Selects Best Lung Cancer Screening Candidates

In the National Lung Screening Trial (NLST), screening for lung cancer with low-dose chest CT scans resulted in a 20% reduction in death from lung cancer. The consumer-serving American Lung Association recommended outright that older people with heavy smoking histories should get lung cancer screening; leading professional societies including the American Thoracic Society (ATS) and the American College of Chest Physicians (ACCP) have more cautiously endorsed lung cancer screening to their physician members.

The ATS and ACCP's caution in part stems from the concern that since the benefit was relatively small (a 20% relative risk reduction) in a high risk group, giving the official thumbs up to lung cancer screening in our screening-obsessed medical care system would likely result in many lower risk people getting screened, in whom the risks of screening (procedures, radiation, anxiety) are likely to outweigh any potential benefits.

But no screening criteria are perfect, and the rub is, some "lower risk" people according to NLST enrollment criteria will still get lung cancer. The NLST only enrolled heavy (30+ pack year) smokers aged 55-74. Everyone else was excluded -- the 73-year old with a 29 pack-year smoking history, the 54-year old with a 60 pack year history and a strong family history of lung cancer -- and they would thus also generally be excluded from screening by the ATS and ACCP's current recommendations. But of course these people are at high risk for lung cancer, too, and might have their lives saved from screening.

Screening guidelines are inherently flawed at accurately discriminating between high and low risk patients, both because risk exists on a continuum and cutoff points are arbitrary, and because in cases where multiple risk factors interact, a guideline's need for simplicity means some risk factors get excluded, along with whatever predictive ability they could have contributed.

Martin C. Tammemägi et al address these issues in a New England Journal of Medicine article proposing a new risk stratification system based on their huge data trove from the Prostate, Lung, Colorectal and Ovarian (PLCO) Cancer Screening Trial (which among other findings, showed that chest X-rays are useless as a screening test for lung cancer).

The idea is, by including a patient's specific risk factors, you can calculate her 6-year risk for lung cancer and make more intelligent screening decisions (by screening higher risk patients, but not the lower risk ones). Their methodology is impenetrable to the non-statistician, but fortunately they've created an Excel spreadsheet for MDummies, which you can access online. (You have to fill out a form first though.)

In addition to age and pack-years smoking history, their risk factors include body mass index, race/ethnicity, educational level, current smoking status, family history of cancer, personal history of cancer, and presence of COPD.

By using the PLCO model's optimal area under the curve, they assert that their model identified 81 more cases of lung cancer than did the NLST, and would have prevented 12 additional deaths. They also found they had fewer missed cancers (false negatives) and fewer false positives.

How? Mainly by identifying high-risk people to screen better than NLST could. As an example, if NLST criteria were applied to the PLCO participants, only 14,144 of 37,332 smokers (38%) would have been eligible for screening. A few cancers would be missed in that ostensibly "lower risk" excluded group. Accuracy was also improved through the robustness of the model and selecting the ideal cutpoint at which to screen, reducing both false positives and false negatives compared to NLST.

Try as I might, I couldn't find the cut point of pretest probability at which they recommend screening for lung cancer, from their paper or their Excel calculator. Turns out, that's because it's not there. I emailed Dr Tammemägi ask his opinion, and he graciously and helpfully gave this response:

If one wants to work in parallel to the NLST one might choose the top 40% of smokers at risk for screening.  This would would mean using a cut-point of >1.27% using the Excel calculator. The NLST criteria leads to screening roughly 38% of smokers.  Screening of the top 40% at model risk (by using a 1.27% cutoff on the calculator) yields a sensitivity for identifying individuals who are diagnosed with lung cancer of 84%, specificity of 60% and PPV of 3.7% in PLCO data.

This provides you with a starting point for recommendations. I am preparing a more detailed analysis exploring the accuracy of screening selection thresholds, and how to interpret them, and will submit the manuscript in the near future.

With further refinement, prediction tools such as that of Dr Tammemägi et al could help harness the benefits of CT screening for lung cancer for the most appropriate patients, reducing the societal risks, costs, and controversy that a blunt nationwide deployment could create. Of course, their findings come from a model; validating it in the real world will take ... well, decades.

I also learned that if you're 80 years old, Native Hawaiian, a current smoker of 2 packs a day for 60 years, with a BMI of 39, COPD, and a history of colon cancer with a family history of lung cancer, and never finished high school, you have a 78% chance of having lung cancer within 6 years. For everyone else, your odds are at least a little better than that.

Martin C. Tammemägi et al. Selection Criteria for Lung-Cancer Screening. N Engl J Med 2013; 368:728-736.

Lung cancer risk calculator for CT screening decisions (Excel spreadsheet download) - requires completion of "captcha" form first.