Long-term lung cancer risk prediction using Sybil AI on routine coronary artery calcium scans: a 15-year study in the Multi-Ethnic Study of Atherosclerosis (MESA) study
M MozafaryBazargany, A Hashemi, A Reeves, K Atlas, C Zhang, H Mohammadi, H Ghoshouni, H Zarei, R Mirjalili, A Azimi, A Ghaffari Jolfayi, Z Fayad, D Yankelevitz, N Wong, M NaghaviAbstract
Background
Sybil is a deep-learning model validated for predicting lung cancer (LC) risk from low-dose CT scans. While typically used for screening-intent imaging, millions of Coronary Artery Calcium (CAC) scans are performed annually for cardiovascular risk assessment. These scans represent an untapped opportunity for opportunistic LC screening. This study evaluates the long-term predictive power of Sybil when applied to baseline CAC scans within a large, multi-ethnic, population-based cohort over a 15-year follow-up period.
Methods
We analyzed 5,726 participants from the Multi-Ethnic Study of Atherosclerosis (MESA Exam 1) who underwent baseline CAC scans (mean age 62.2±10.3 years). Sybil’s predictive performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC) for LC incidence at intervals from 1 to 15 years. The model assessed future malignancy using only baseline non-contrast, cardiac-gated images, without clinical risk factors or smoking history.
Results
In the MESA cohort, 45 participants developed LC within 6 years. Sybil demonstrated strong short-term and sustained long-term predictive accuracy using only baseline CAC scans. The AUC for 1-year LC prediction was 0.81, remained high at 0.80 for 2-year prediction, and stabilized at 0.75 for 6-year risk. Remarkably, the model maintained significant predictive power for up to 15 years, with an AUC of 0.62. In the 6-year LC group, the mean Sybil output score was significantly higher compared to non-cancer participants (18.0% vs. 6.2%, p<0.001). No statistically significant degradation in predictive performance was observed over the mid-term follow-up, suggesting the model captures systemic pulmonary risk signals rather than relying solely on localized lesion detection in the truncated cardiac field of view.
Conclusion
Sybil AI provides robust, sustained lung cancer risk prediction for up to 15 years using a single baseline CAC scan. Integrating Sybil into routine cardiac imaging can identify high-risk individuals years before clinical manifestation, transforming CAC scans into a multidimensional prevention platform.time dependent ROC curve