麻豆传媒

CHICAGO -- An artificial intelligence (AI) algorithm successfully identified never smokers at high risk for lung cancer using existing x-rays in the electronic medical record (EMR), a researcher reported here.

Among a cohort of more than 17,000 documented never smokers with a chest x-ray originally obtained for cough, fever, or some other routine indication, the deep-learning model deemed 28% to be high risk, 58% at moderate risk, and 13% at low risk for lung cancer.

Over 6 years from the time of the x-ray, incident lung cancer was detected in 2.9%, 1.8%, and 1.1% of these groups, respectively, according to Anika Walia, BA, a medical student at Boston University, who presented the findings at the here.

That 2.9% incidence in the high-risk group well exceeds the 6-year risk threshold (>1.3%) where low-dose CT-based lung cancer screening is recommended by guidelines, she said.

Lung cancer is becoming more common in never smokers (roughly 10% to 20% of new cases), a group that often presents with advanced-stage disease, but current Medicare and U.S. Preventive Services Task Force guidelines only recommend low-dose CT screening for individuals with a significant smoking history.

"Since cigarette smoking rates are declining, approaches to detect lung cancer early in those who do not smoke are going to be increasingly important," co-investigator Michael Lu, MD, MPH, of Massachusetts General Hospital in Boston, said in a press release.

The study used the CXR-Lung-Risk algorithm, which was trained on chest x-rays from more than 40,000 smokers and never smokers in the Prostate, Lung, Colorectal, and Ovarian () cancer screening trial.

Walia showed that the algorithm turned in similar -- or the degree of agreement between two measuring or rating techniques -- even when other factors such as demographics (age, sex, race) or medical history (history of chronic obstructive pulmonary disease [COPD], asthma, or lower respiratory tract infection) were taken into consideration:

• CXR-Lung-Risk alone: C-statistic 0.61 (95% CI 0.58-0.64)
• Demographics alone: C-statistic 0.61 (95% CI 0.58-0.64)
• Demographics plus medical history (full model): C-statistic 0.62 (95% CI 0.59-0.65)
• CXR-Lung-Risk plus full model: C-statistic 0.64 (95% CI 0.61-0.67)

"There is information on the chest x-ray about the individual's health and risk of cancer that we currently don't use but can be extracted from the image using AI, as shown by the concordance statistics," Walia told 麻豆传媒.

Walia acknowledged that the C-statistic may be too low for the deep-learning model to be used as a screening instrument, but suggested the tool could help identify high-risk nonsmokers for lung screening trials.

Mylene Truong, MD, of the University of Texas MD Anderson Cancer Center in Houston, told 麻豆传媒 that "more work needs to be done before this could be used generally," but agreed the algorithm could be helpful for finding high-risk patients to enroll for low-dose CT screening.

"This may lead to changes to eligibility criteria for lung cancer screening, specifically to include nonsmokers at high risk," said Truong, who was not involved in the study.

But she noted study limitations as well, including that information on patients' family history -- a risk factor for lung cancer -- was not available, nor was it clear what characteristics were being analyzed on the x-ray exactly.

The study from Walia's team included 17,407 never smokers (mean age 63 years, 57% women) with a routine chest x-ray obtained from 2013 to 2014, including 4,860 deemed high risk by the algorithm, 10,031 moderate risk, and 2,176 low risk. Those with a previous lung cancer diagnosis, a history of screening for lung cancer, or poor quality x-rays were excluded.

A majority were white (60%), with 7% Black, 5% Asian, and 4% of Hispanic ethnicity. COPD was documented in 5%, asthma in 14%, and a prior lower respiratory tract illness in 19%. Overall, lung cancer was detected in 2% of the population in the 6 years after the initial x-ray.

The researchers noted that validation of the algorithm in a diverse patient population will be needed.

Comment