Performance and Utility of the Sybil Deep Learning Model for Lung Cancer Risk Prediction in Asian High- and Low-Risk Populations.
Authors
Affiliations (7)
Affiliations (7)
- Division of Pulmonary and Critical Care Medicine, Department of Internal Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea; Department of Internal Medicine, Seoul National University College of Medicine, Seoul, Republic or Korea. Electronic address: [email protected].
- Seoul National University College of Medicine, Seoul, Republic of Korea; Center for Artificial Intelligence in Healthcare, Seoul National University Bundang Hospital Healthcare Innovation Park, Seongnam, Republic of Korea.
- Center for Artificial Intelligence in Healthcare, Seoul National University Bundang Hospital Healthcare Innovation Park, Seongnam, Republic of Korea.
- Big Data Center, Seoul National University Bundang Hospital, Seongnam, Republic of Korea.
- Department of Radiology, Seoul National University Bundang Hospital, Seongnam, Republic of Korea; Department of Radiology, Seoul National University College of Medicine, Seoul, Republic of Korea.
- Division of Pulmonary and Critical Care Medicine, Department of Internal Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea; Department of Internal Medicine, Seoul National University College of Medicine, Seoul, Republic or Korea.
- Division of Pulmonary and Critical Care Medicine, Department of Internal Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea; Emeritus Professor, Seoul National University College of Medicine, Seoul, Republic of Korea.
Abstract
Sybil is a deep learning model designed to predict future lung cancer risk based on a single low-dose chest CT (LDCT) scan, facilitating a precision-based approach to lung cancer screening (LCS). What are the performance and utility of Sybil in a pragmatic Asian LCS cohort? We analyzed 21,087 individuals aged 50-80 who underwent LDCT screening between January 2009 and December 2021. Baseline LDCT scans were evaluated using Sybil to calculate the risk of lung cancer diagnosis within 1 to 6 years. Model performance was assessed with the area under the receiver operating characteristic curve (AUROC). Stratified analyses were conducted for individuals with ≥20 pack-years (PY) of smoking history (n=4,611), ever-smokers with <20 or unknown PY (n=5,378), and individuals who have never smoked (INS) (n=11,098). Among 21,087 participants, subsolid nodules were more common in INS (9.3%) than in those with ≥20 PY (7.1%) or <20/unknown PY (6.7%) (p<0.001). A total of 257 participants were diagnosed with lung cancer within 6 years. Sybil achieved AUROCs of 0.86 (95% CI=0.82-0.89) for 1-year and 0.74 (95% CI=0.70-0.77) for 6-year invasive lung cancer predictions. For INS, AUROCs were 0.86 (95% CI=0.81-0.91) for 1-year and 0.79 (95% CI=0.74-0.84) for 6-year predictions. Sybil's short-term performance declined in individuals with LDCT findings of granulomatous sequelae, but long-term performance was maintained. Reduced performance was also observed in individuals with subsolid nodules and those without a detected baseline nodule. Five-year overall survival among lung cancer cases was 92.7% in INS, 75.9% for <20/unknown PY, and 68.6% in ≥20 PY (p<0.001). Sybil demonstrated robust performance in predicting future lung cancer in an Asian screening cohort comprised of individuals with diverse risk profiles. The findings highlight the potential to develop personalized LCS strategies using Sybil to address region-specific needs.