Long the domain of astrologers and tarot card readers, prediction has recently become downright fashionable. While quant-minded individuals like Billy Beane and Nate Silver have achieved fame and fortune using probabilistic forecasting, dozens of smartphone apps deliver the predictive insight of clinical risk scores to doctors’ fingertips. Why all the enthusiasm? Accurate predictions allow us to prepare for the future.
Testing their predictive mettle in this week’s NEJM, Dr. Annette McWilliams (British Columbia Cancer Agency, Vancouver, Canada) and colleagues ask a deceptively simple research question: If a low-dose computed tomography (LDCT) lung cancer screening test detects a lung nodule, can we use the information at hand to accurately predict if it is malignant?
Using clinical and LDCT data from 1871 current or former smokers in the PanCan study, the investigators developed a model to predict when a newly discovered nodule was cancerous. Model variables included age, family history of lung cancer, and the presence of emphysema as well as nodule size, type, and location. Next, the investigators tested this prediction model in a cohort of 1090 current and former smokers enrolled in several British Columbia Cancer Agency chemoprevention trials. They found their model successfully discriminated between higher-risk and lower-risk nodules even within this validation cohort (AUC = 0.97, 95%CI 0.95-0.99), suggesting that the model can also be generalized to other groups of current and former smokers.
How might this be clinically useful? Let’s say you are seeing a former smoker with a nodule discovered on an initial LDCT screening test. A “low risk” designation on the predictive model suggests that you can reassure your patient that the nodule is very unlikely to be cancerous (with a negative predictive value of 99.6%). As a result, the two of you might decide that a less aggressive diagnostic strategy might be most appropriate.
“LDCT screening has been shown to reduce lung cancer mortality,” says hematologist-oncologist and NEJM Deputy Editor Dr. Dan Longo, “but the optimal management strategy for lung nodules detected by screening hasn’t been established. Accurate risk stratification might help guide decisions such as the frequency of repeat imaging or the necessity of a more invasive test.”
How exactly will this risk stratification tool influence lung cancer screening guidelines? Predicting that kind of outcome is difficult, especially since we don’t have a regression model to help. Tarot cards, anyone?
View the latest NEJM Quick Take, an animated overview of the study.