Home > Oncology > Neural network identifies high-risk smokers for lung-cancer screening

Neural network identifies high-risk smokers for lung-cancer screening

Annals of Internal Medicine
Reuters Health - 01/09/2020 - A new neural network can use chest x-rays to improve the identification of high-risk smokers for lung-cancer screening CT, researchers report.

"We designed this model to take inputs that are commonly available in day-to-day practice - a chest x-ray image, age, sex, and whether the person is a current smoker," explained Dr. Michael T. Lu of Massachusetts General Hospital and Harvard Medical School, in Boston.

"This is different than the way we assess lung-cancer risk now, which is based on the amount of smoking (e.g., number of cigarette packs per day times the number of years of smoking)," he told Reuters Health by email.

"Limitations of this smoking-based approach are that people may not recall their smoking history accurately and that there is variability in how peoples' lungs respond to smoking," he said.

Dr. Lu and colleagues developed their model, called CXR-LC, using data from the PLCO Cancer Screening Trial and validated it in 5,615 PLCO smokers and 5,493 heavy smokers from the National Lung Screening Trial (NLST). They also compared the model's performance with Centers for Medicare and Medicaid Services (CMS) eligibility and with the Modified PLCO Lung-Cancer Risk-Prediction Model (PLCO-M2012) score.

In the PLCO validation data set, CXR-LC was significantly better at discriminating between persons who did and did not develop lung cancer compared with CMS eligibility (AUC, 0.755 vs. 0.634). But it had similar discrimination to PLCO-M2012 (AUC, 0.761), the researchers report in the Annals of Internal Medicine.

Among CMS-eligible smokers, discrimination was similar for CXR-LC and PLCO-M2012 in both the PLCO and NLST validation sets.

Adding CXR-LC to PLCO-M2012 yielded a modest improvement in discrimination in both validation sets.

When CXR-LC scores were divided into three risk groups, lung-cancer incidence rates per 1,000 person-years for very high versus very low CXR-LC risk were 12.4 versus 1.1 for PLCO and 12.7 versus 2.3 for NLST.

While CXR-LC was developed to predict incident lung cancer, it also predicted 12-year lung-cancer mortality with greater accuracy than CMS eligibility and with accuracy similar to that of PLCO-M2012.

In subgroup analyses, CXR-LC had greater benefit than the other two models in men, whereas in women CXR-LC had net benefits lower than PLCO-M2012 but similar to CMS eligibility.

"Although the cost ($21 in CMS and £25 in the U.K. National Health Service) and radiation dose (equivalent to 10 days' natural background radiation) are low, we would not recommend chest radiography solely to assess lung cancer risk," the authors note. "Instead, a pragmatic future implementation of CXR-LC could analyze existing chest radiographs from outpatient smokers by using an automated electronic medical records (EMR) tool."

"My hope is that these technologies will help us better identify people at high-risk of cancer, but also save physicians time and attention that they can redirect towards their patients," Dr. Lu said. "This is needed because currently <5% of high-risk smokers have lung-cancer screening CT. Getting more of these high-risk people into the screening pipeline would prevent lung cancer deaths."

By Will Boggs MD

SOURCE: https://bit.ly/3gN9r15 Annals of Internal Medicine, online September 1, 2020.

Posted on