Introduction
Lung cancer screening by low-dose CT has been shown to reduce lung cancer mortality for people at high risk based on age and smoking history. However, for people at low risk, harms of screening might outweigh benefits. Implementation of lung cancer screening is progressing in many countries. Most countries define eligibility for screening using ‘categorical’ criteria. Alternatively, eligibility may be defined by risk prediction models that estimate an individual’s probability of being diagnosed with lung cancer based on their detailed risk factors. However, many models have been developed, and they may perform differently across geographical regions or ethnic groups, but haven’t been comprehensively validated. This abstract introduces 5 studies that we conducted to validate the performance of up to 16 lung cancer prediction models across geographical contexts and ethnic groups.
 
Method
We analyzed 24 population cohorts from Asia, Europe and North America in the harmonized database the Lung Cancer Cohort Consortium (LC3). We included  individuals with a history of smoking and aged 40-80 years old. We quantified the performance of prediction models as calibration and discrimination. We compared the eligibility and sensitivity for each model in each population with USPSTF-2021 criteria. In France and Brazil, we have modeled the potential impact of different strategies for eligibility on national lung cancer screening.
 
Results
Well-performing lung cancer risk models have been developed using data largely from individuals living in North America or Europe. Among individuals from Asia, the models do not perform as well as people from these regions.  Similarly, model performance is reduced among minority racial and ethnic groups among individuals in the United States.
 
Applying USPSTF-2021 among population aged 50-80 with a smoking history identified 25%-40% of them as screening eligible and 65%-75% of future lung cancer cases within 5 years would be identified. Among the many lung cancer risk prediction models that have been developed, a small group of models offers similarly better performance compared to USPSTF-2021 criteria, and would be reasonable to use when selecting individuals for lung cancer screening. Taking France as an example, where a national lung cancer pilot programme is being launched, use of a well-performing risk prediction model for eligibility could identify an additional 12-22% of lung cancer deaths compared with categorical criteria. Modeling study for Brazil showed similar results.
 
Conclusion
Lung cancer risk prediction models can more efficiently identify people who are likely to benefit from screening compared to categorical criteria. Lung cancer screening studies, pilots, and programmes are encouraged to consider using a risk-based approach to selecting eligible individuals, or to evaluate such an approach alongside other considered criteria.