External Validation of Lung Cancer Prediction Models Combining Epidemiological Predictors in Chinese Ever and Never Smokers: Guangzhou Biobank Cohort Study.

Objective: This study aimed to externally validate existing lung cancer models using data from the Guangzhou Biobank Cohort Study (GBCS) and compare their predictive performance for Chinese ever and never smokers.

Methods: We evaluated the discrimination and calibration of LCRAT (Lung Cancer Risk Assessment Tool), LLP version 2 (Liverpool Lung Project version 2), LLP version 3 (Liverpool Lung Project version 3), HUNT (HUNT was derived from the Nord-Trøndelag Health Study), OWL (Optimized Early Warning Model for Lung Cancer Risk), LCRS (Lung Cancer Risk Score), PLCOm2012 (Prostate, Lung, Colorectal, and Ovarian 2012 model), PLCOall2014 (Prostate, Lung, Colorectal, and Ovarian 2014 model), NHIS (Korean National Health Insurance Service), LLPi (Liverpool Lung Project Risk Prediction Model for Lung Cancer Incidence), Pittsburgh, and Bach models. We compared the performance of models and Chinese lung cancer screening (T/CPMA 013-2020), US Preventive Services Task Force 2021 (USPSTF-2021) and Nederlands-Leuvens Longkanker Screenings Onderzoek (NELSON) criteria.

Results: The LLP version 2, LLP version 3, OWL, LCRS, PLCOall2014, and LLPi models showed better performance in ever smokers than in never smokers, with higher AUC (0.72-0.82 vs. 0.69-0.71) and E/O (expected to observed) ratios (0.57-0.79 vs. 0.60-0.69), while the LCRAT, HUNT, PLCOm2012, NHIS, Pittsburgh, and Bach models showed good performance in ever smokers, with AUC ranging from 0.70 to 0.79 and E/O ratios from 0.57 to 0.75. The T/CPMA 013-2020, USPSTF-2021, and NELSON criteria identified 56.52%-75.58% of high-risk individuals at 5, 6, 6.6, 8.7, and 10 years, while the LCRAT, LLP version 2, LLP version 3, HUNT, OWL, LCRS, PLCOm2012, PLCOall2014, NHIS, LLPi, Pittsburgh, and Bach models identified 70.70%-89.72% of high-risk individuals.

Conclusions: Most lung cancer risk prediction models showed good performance and identified more cases than screening criteria. Replacing screening criteria with risk prediction models may increase lung cancer screening efficiency.