How to ration lung cancer screening CTs, rationally
Restrict Lung Cancer Screening CT To Highest-Risk People?
The National Lung Screening Trial showed that 3 annual low-dose chest CT scans in people with heavy smoking histories (30+ pack-years) saved lives, reducing the risk of death from lung cancer by a relative 20% compared to screening with annual chest X-rays. With 160,000 lung cancer deaths in the U.S. each year, widespread CT screening could theoretically prevent 12,000 lung cancer deaths each year.
Skeptics pointed out that the absolute risk reduction in lung cancer death seen in NLST was rather low, requiring 320 people to be screened with CT in order to prevent one death from lung cancer. Along the way, more than one-third of the ~27,000 screened had false positive CT scans, with associated health care expenses, anxiety, and avoidable biopsies. Widespread lung cancer CT screening of the 9 million U.S. adults meeting NLST criteria has been estimated to cost $1.5 billion / year and about $240,000 per life saved. The leading expert bodies in pulmonary medicine cautiously endorsed lung cancer screening, but only for people meeting entry criteria for the NLST (age 55-74, 30+ pack-years, quit < 15 years ago).
The subjects enrolled in the National Lung Screening Trial were all at elevated risk for lung cancer, but even among them, there was a wide spectrum of risk. Prior investigators have calculated a 10-fold difference in the benefit of lung cancer screening (i.e., number of lung cancer deaths prevented) between those in the highest and lowest risk strata in NLST. (Here's a calculator to help stratify lung cancer risk for your patient.)
Stephanie Kovalchik et al sliced, diced, and chopped the NLST data to help us better understand this spectrum of lung cancer risk within the heavy-smoking-history population, providing a basis for intelligent and potentially life-saving individualized recommendations for lung cancer screening to our patients. Their results are in the July 18, 2013 New England Journal of Medicine.
What They Did
They identified factors within NLST participants known to be associated with death from lung cancer -- age, pack-years and years since quit, BMI, emphysema, and family history of lung cancer, and created an a priori prediction model based on these variables. They externally validated this lung cancer risk calculator using the data set from another lung cancer screening trial (PLCO). They then divided the NLST population into five equal quintiles of lung cancer death risk (~5,300 in each group), and inserted the NLST outcomes data to analyze the efficacy of lung cancer CT screening according to the pretest model-calculated risk. Pretest model-calculated 5-year lung cancer death risk ranged from 0.15 - 0.55% in the lowest cohort (quintile 1) to >2.0% in quintile 5.
What They Found
Those in the lowest a priori or pretest risk quintile 1 got the lowest benefit of CT screening for lung cancer. They had screening CTs that were most likely to "cry wolf", with a high proportion of false positive lung cancer screening CTs and low probability of identifying curable lung cancer. The incidence of lung cancer in this segment was only about 1.5%, with a death rate of 0.4% -- and only 1 life was saved by CT screening, from over 5,200 people screened.
At the other extreme, people in quintile 5 had an almost 10% chance of developing lung cancer during the trial, and a 3% of dying from lung cancer during follow-up. CT screening saved 33 lives in this highest-risk group, with a number needed to screen of 161 to save a life.
Quintile 4 (the second-highest) was very similar to quintile 5 (number needed to screen to prevent a lung cancer death: 171), making this top 40% block an apparently better candidate group for lung cancer screening. Quintiles 2 and 3's number needed to screen were 415 and 531.
What It Means
Authors estimate that by screening only people meeting NLST criteria in the top 3 quintiles of risk (60% of NLST-eligible people) would dramatically improve the performance of CT screening for lung cancer:
Almost 90% of people with curable lung cancer could be identified.
The number needed to screen would be reduced from >300 to 161.
False positive screening CTs would be cut from >100 to only 65 for every prevented lung cancer death.
A relatively small proportion of people with curable lung cancer (~13%) would be missed by lack of screening.
Now, how can we use this in practice? Can we identify people in a clinic setting who are at the highest risk for lung cancer, corresponding to the higher quintiles in this analysis?
The short answer is, no, not by using the data from this paper. The authors don't share their secret formula for calculating a patient's a priori risk in the paper, but perhaps they will (I've emailed them to ask). You can't really "eyeball" their published risk factors and hazard ratios either (Table 2), because most are on the order of 1.1 - 1.5 and all of them together only change most patients' absolute 5-year risk of death from lung cancer by 0.5% or so. The power of their prediction model derives from the large sample size.
You can use this lung cancer risk calculator, based on a similar philosophical approach and published in NEJM earlier this year. As they're developed and refined, perhaps such calculators will become available online in a more user-friendly format, where they can be used by physicians at the point of care to help best harness the potentially lifesaving power of CT screening for lung cancer.
Update: Stephanie Kovalchik, lead author of the paper was kind enough to share some additional information on their study, including why they didn't share their risk model straightaway, and what we can expect from their group in the future. Her emailed comments:
By considering more detailed information about an individual's smoking history (pack-years, quit-years) and demographic/clinical characteristics (age, race, family history, emphysema, etc.), our prediction model is much more specific than NLST entry criteria. Our study showed that more detailed risk information than NLST entry criteria is important for determining the expected benefit a smoker is likely to receive with CT screening. Although this has clear implications for using risk prediction models to target screening, before a risk model is used for screening recommendations, it should be validated in a general population of smokers (as opposed to a research cohort, which may not be exactly representative of smokers of the same age in the general population). We are currently working on a validation study for the lung-cancer death prediction model in a nationally-representative cohort of US smokers. This study will clarify the applicability of the model to US smokers. We plan to publish our findings and guidelines for the use of the lung-cancer death prediction model with an accompanying web tool in the near future.