Lung cancer risk in never-smokers predicted by AI tool 'Sybil'

Yahoo

19-05-2025

ST. PAUL, Minn., May 19 (UPI) -- With lung cancer rates among non-smokers rising, especially young East Asian women, a new study released Monday is touting the promise of an artificial intelligence tool to "strongly" predict who's most at risk.
Lung cancer has long been associated with smoking. But even as overall rates steadily drop and smoking decreases around the world, a unique population of young East Asians are seeing a 2% annual increase in lung cancer cases -- even though half of them have never smoked.
The cause of this remains unknown, but suspicion is centered on genetic mutations developed during a person's lifetime rather than inherited, such as damage to a gene that codes for a protein known as EGFR, which prevents cells from growing too quickly.
This genetic damage is believed to be caused by environmental toxins including second-hand smoke and even fumes produced by high-temperature stir-fry cooking in rooms that lack proper ventilation.
Globally, more than 50% of women diagnosed with lung cancer are non-smokers, compared to 15% to 20% of men. Meanwhile, an estimated 57% of Asian-American women diagnosed with lung cancer have never smoked, compared to only about 15% of all other women, according to a recent University of California-San Francisco study.
Against this backdrop of rising cancer cases among seemingly low-risk women, the potential of AI to accurately predict who may be most suspectable to a surprise lung cancer diagnosis has generated considerable interest around the world.
In a paper presented Monday at the American Thoracic Society's medical conference in San Francisco, Dr. Yeon Wook Kim of the Seoul National University Bundang Hospital reported a new AI tool dubbed "Sybil" has proven to be accurate in identifying which "true low-risk individuals" are more likely to develop lung cancer -- all foretold from a single low-dose chest CT scan, or LDCT.
Sybil, named after the female seers of ancient Greek mythology, was developed in 2023 by researchers at the Massachusetts Institute of Technology's Abdul Latif Jameel Clinic for Machine Learning in Health, the Mass General Cancer Center and Chang Gung Memorial Hospital in Taiwan.
It was trained first by feeding it LDCT images largely absent of any signs of cancer, since early-stage lung cancer occupies only tiny portions of the lung and is invisible to the human eye. Then, researchers gave Sybil hundreds of scans with visible cancerous tumors.
In its first run, Sybil was able to deliver "C-indices" of up to 0.81 in predicted future occurrences of lung cancer from analyzing one LDCT. Models achieving predictive C-index scores of over 0.7 are considered "good" and those over 0.8 are "strong."
This week's Korean study validated those results. Kim and his colleagues evaluated 21,087 people ages 50 to 80 who underwent self-initiated LDCT screening between January 2009 and December 2021 in a tertiary hospital-affiliated screening center in South Korea. These subjects were followed up until June 2024.
Baseline LDCTs were analyzed with Sybil to calculate the risk of lung cancer diagnosis within one to six years. Analyses were performed for individuals with various smoking histories, ranging from more than 20 "pack-years" to never-smokers, who comprised 11,098 of the participants.
Among all participants, 257 (including 115 never-smokers) were diagnosed with lung cancer within six years from the baseline LDCT. Sybil achieved a C-index for lung cancer prediction at one year of 0.86 and 6 years of 0.74 for all the participants, while among never-smokers, one-year and six-year C-indices were 0.86 and 0.79, respectively.
Kim told UPI the results hold the promise of helping to regularize lung cancer screening in Asia, where those efforts are inconsistent and, due to differing demographics, sometimes are at a "disconnect" with international screening criteria.
"Asia bears the highest burden of lung cancer, accounting for over 60% of new cases and related deaths worldwide," he said in emailed comments. "A growing proportion of this burden is observed among individuals who have never smoked, particularly among women.
"In Korea, more than 85% of female lung cancer patients are non-smokers. As a result, increasing attention has been given to evaluating the effectiveness of lung cancer screening, or LCS, in traditionally low-risk populations in Asia."
Government-led programs and initiatives have expanded to include never-smokers into their LCS efforts, while other efforts varying from international guidelines due to their inclusion of such never-smokers have "gained traction in East Asian countries, including South Korea, Taiwan and China," Kim said.
AI tools like Sybil could be used to develop "personalized strategies" for patients who have already undergone LDCT screening, but have not yet had follow-ups, he added, while cautioning that further validation will be needed "to confirm the model's potential for clinical use.
"While the need for screening low-risk groups may be justified in certain settings, the lack of evidence from randomized trials limits the development of long-term LCS strategies for these populations."
Researchers, meanwhile, are "actively" working on expanding Sybil's uses into other personalized health applications, said Adam Yala, an assistant professor at the UCSF/UC-Berkeley Joint Program in Computational Precision Health and one of the AI model's developers.
"One, this is broadly applicable across many different types of cancers," he told UPI. "We've got processes ongoing for breast cancer, and we're also working on prostate and pancreas cancers.
"And there's also evidence that from CT scans you could predict sudden deaths from cardiovascular disease. This would provide early detection, giving you a better opportunity for early intervention to provide better outcomes. So it's not uniquely about cancer. ... There's a version of this for cardiovascular health, and there could be other areas of medicine, as well."
AI's potential to provide health benefits, Yala added, "is totally untapped. For instance, now we're only looking at a patient's CT scan once, but over time, you could look at multiple CTs. Mammograms, as well. There's a lot of data available there. It's a field at its infancy."