Lung cancer remains the leading cause of cancer deaths worldwide, despite significant therapeutic advances, improvements in early detection methods, and the implementation of large-scale screening of individuals with a history of smoking tobacco. But screening gaps and disparities persist.
“A significant unmet need in pulmonary medicine as related to cancer is that lung cancer kills more people than colon, prostate, and breast cancer combined,” said William Mayfield, MD, Medical Director of Lung Cancer Screening and Incidental Nodule Programs at Wellstar Health System in Georgia. Dr. Mayfield is a principal investigator in the Sybil Implementation Consortium, an alliance focused on the development and implementation of the Sybil artificial intelligence (AI) model for predicting lung cancer risk.
Despite the mortality burden, only about 20% of the eligible individuals are screened for lung cancer.1
Sybil Implementation Consortium member Mary Pasquinelli, DNP, APRN, FNP-BC, Director of the Lung Screening and Early Detection Program at the University of Illinois Chicago, said that, currently, lung cancer screening is covered by insurance under the United States Preventive Services Task Force guidelines.
“Unfortunately, this screening approach misses about 50% of the people eventually diagnosed with lung cancer,” she said.
Dr. Pasquinelli explained that current guidelines recommend lung cancer screening based primarily on smoking history. However, she said, about 20% of lung cancers globally occur in people who have never smoked before, and lung cancer in this population is the fifth leading cause of global cancer deaths.
“Although other factors also contribute to lung cancer risk, the cumulative impact of lifelong insult on lungs—from not just tobacco exposure but also air pollution, respiratory diseases, and the interplay between biology, environment, and social determinants of health—[is] not fully understood,” Dr. Pasquinelli said.
Improving lung cancer risk prediction
Risk models that identify candidates for screening based on factors beyond age and smoking history and that leverage clinical, demographic, and/or imaging data, including AI-enabled models, have gained attention in this context. Dr. Pasquinelli noted the example of PLCOm2012, a regression lung cancer risk prediction model based on 11 risk factors applicable in diverse populations.2
While some risk models demonstrate robust predictive performance, they serve only to identify individuals for screening.3 With increasing deployment of imaging and an expansion of AI applications in radiology, several AI-based risk models have been developed that leverage epidemiologic factors, along with digitized imaging data and/or clinical parameters.2
The prospective observational Detecting Early Lung Cancer (DELUGE) study demonstrated higher lung cancer detection rates in an incidental nodule program, which included patients who may not have otherwise qualified for screening.4
“This finding set off a national firestorm, wherein the central question was whether incidental lung nodule-related data present an opportunity to save lives by improving lung cancer detection using an AI-enabled model for evaluating chest radiology images,” Dr. Mayfield said.
Florian Fintelmann, MD, Associate Professor of Radiology, Harvard Medical School, said there already are several AI-enabled tools, which he and colleagues reviewed in a recent series on opportunistic screening.5
“What these tools are lacking is that they don’t provide a global future lung cancer risk prediction. These tools are nodule-centric,” he said.
Nodule-centric risk prediction models may miss information beyond nodule features that is contained in low-dose chest CT (LDCT) scan images and that may help predict future lung cancer risk, he explained.
Enter Sybil: an AI model that predicts an individual’s future lung cancer risk from a single LDCT scan.
Design and development of Sybil
Dr. Mayfield described how Sybil was developed by training an AI algorithm on 44,000 LDCT scans from the National Lung Screening Trial (NLST).6
“The intent was to use any LDCT scan and derive a predictive score for the risk of developing lung cancer over the subsequent six years, without any other clinical information or medical or family history.”
Dr. Fintelmann said Sybil was developed using a six-year window for risk prediction because current risk predictors, such as PLCOm2012, also use this window.
“When we developed Sybil, there was no real comparator against which to benchmark the model; unlike other models, Sybil relies solely on an LDCT [scan] image for risk prediction,” he said. “Sybil does not rely on clinical information. This was a deliberate design choice because substantial evidence and our personal clinical experience indicate that obtaining clinical information, including family history and patient exposures, is time-intensive and is often not reliably captured in electronic medical records. Sybil provides a lung cancer risk prediction solution that is independent of known and unknown risk factors, improving convenience and saving time.”
Implementing Sybil
Sybil has been deployed in 25 hospitals across 11 countries and validated in more than 120,000 LDCT scans. This success has not been without challenges though.
“One of the unexpected challenges was getting Sybil, an open-source [and] freely available software that lives on a GitHub repository, into the hands of providers. Paradoxically, health care systems find it harder to integrate tools that are free and/or not backed by a commercial vendor,” Dr. Fintelmann said.
Indeed, implementing Sybil at Wellstar Health “was a heavy lift that took over a year,” Dr. Mayfield noted.
Dr. Fintelmann said that one of the immediate deliverables that the Sybil Implementation Consortium is working on is “a playbook for Sybil implementation.”
Validating Sybil in broader populations
“Sybil was initially validated in the NLST, which included 91% White patients and only 4.5% Black patients. It was, therefore, critical for this AI-enabled model to be validated in different populations,” Dr. Pasquinelli noted.
In her presentation at the International Association for the Study of Lung Cancer 2025 World Conference on Lung Cancer, Dr. Pasquinelli shared data validating Sybil in a cohort of predominantly Black individuals.7 Sybil has also been validated in an East Asian population, which included people with and without a history of smoking.
“The data suggest that Sybil is race- and ethnicity-agnostic, which is wonderful,” Dr. Pasquinelli said.
The Sybil Implementation Consortium is planning to further evaluate Sybil in the context of two prospective trials, including a trial with a cohort with incidental pulmonary nodules, representing a broad population of individuals in terms of their ages and smoking histories.
Opportunities and challenges ahead
At present, Sybil remains a research tool; however, Dr. Fintelmann expressed optimism for the future of Sybil and other AI models.
“I am looking forward to the future,” he said. “We are developing these tools with the goal of improving patient care.”
Dr. Pasquinelli identified three domains of lung cancer care that Sybil may improve in the future:
- Sybil-based risk stratification may help personalize screening intervals.
- As a risk model that does not rely on smoking history, Sybil may help address the stigma associated with a lung cancer diagnosis.
- Perhaps, in the far future, a model like Sybil could help “prediagnose” lung cancer, potentially identifying individuals who may benefit from early therapeutic intervention.
“Even if Sybil is not the model that is eventually used routinely, this field is evolving rapidly” Dr. Fintelmann said. “As we develop Sybil further, we are building the understanding, infrastructure, and pathways necessary to integrate such models into patient care.”
References
1. M. Pasquinelli M, C.H. Durney CH, T. Hill T, et al. MA05.03 validation of Sybil in a minority population: results from the Sybil Implementation Consortium: UIC, MGH, Baptist, Wellstar. J Thorac Oncol. 2025;20(10)(Suppl 1):S72. doi: 10.1016/j.jtho.2025.09.132
2. Latest ACS lung cancer data: only 1 in 5 eligible adults in U.S. screened for lung cancer; 62,000 lives over 5 years could be saved if all eligible screened. American Cancer Society. https://pressroom.cancer.org/2025-lung-cancer-data
3. Pasquinelli MM, Tammemägi MC, Kovitz KL, et al. Risk prediction model versus United States Preventive Services Task Force lung cancer screening eligibility criteria: reducing race disparities. J Thorac Oncol. 2020;15(11):1738-1747. doi:10.1016/j.jtho.2020.08.006
4. Leonard S, Patel MA, Zhou Z, et al. Comparing artificial intelligence and traditional regression models in lung cancer risk prediction using a systematic review and meta-analysis. J Am Coll Radiol. 2025;22(6):675-690. doi:10.1016/j.jacr.2025.02.042
5. Osarogiagbon RU, Liao W, Faris NR, et al. Lung cancer diagnosed through screening, lung nodule, and neither program: a prospective observational study of the Detecting Early Lung Cancer (DELUGE) in the Mississippi Delta cohort. J Clin Oncol. 2022;40(19):2094-2105. doi:10.1200/JCO.21.02496
6. Thuere KL, Mantz L, Sultana S, et al. Opportunistic screening on chest CT, from the AJR Special Series on Screening. AJR Am J Roentgenol. Preprint. Published online July 16, 2025. doi:10.2214/AJR.25.33069
7. Mikhael PG, Wohlwend J, Yala A, et al. Sybil: a validated deep learning model to predict future lung cancer risk from a single low-dose chest computed tomography. J Clin Oncol. 2023;41(12):2191-2200. doi:10.1200/JCO.22.01345
