Evaluation of lung cancer risk prediction models to select lung cancer screening participants in China: a real-world analysis in regional healthcare big data, Yinzhou, China

Background

A number of lung cancer prediction models have been developed worldwide. However, few validation studies have been conducted on Chinese populations. The objective of this study is to evaluate the feasibility and efficacy of 17 global lung cancer risk prediction models when applied to Chinese healthcare big data.

Methods

The study included individuals with information recorded in the Yinzhou regional health care database from January 1, 2010 to December 31, 2021. Seventeen lung cancer risk prediction models (Bach, Spitz, Hoggart, PLCO_m2012, Korean Men, PLCO_all2014, Pittsburgh Predictor, LLPi, LCRAT, HUNT, JPHC, Reduced HUNT, LLPv3, LCRS, OWL, UCL-I, Shanghai-LCM) were evaluated for their performance in overall population and subgroups. The discrimination of the 17 models was assessed using the Harrell's C-index and time-dependent area under the curve (AUC) as metrics. The calibration of the models was evaluated using the expected-to-observed ratio (EOR) and calibration curves. Moreover, the models were recalibrated in the Yinzhou population, and the calibration of the recalibrated models was evaluated.

Findings

A total of 907,200 study participants were included in the analysis, comprising 69,263 smokers and 837,937 non-smokers. Of the 17 models initially considered, only 6 (Bach, Hoggart, Pittsburgh Predictor, JPHC, Reduced HUNT, UCL-I) were available in the Yinzhou regional health care database with complete predictor data. Models that predicted risk over a ten-year period (Bach, JPHC, LCRS, and Shanghai-LCM) exhibited C-indices and AUCs of 0.75 or greater in the ever smokers. The majority of models demonstrated an overestimation of incidence risk in the ever smokers and an underestimation in the never smokers. The JPHC and LCRS models exhibited the most optimal calibration curves and the best EOR, whereas the other prediction models had suboptimal calibration. After recalibration, all models showed improved calibration; meanwhile, the JPHC and LCRS models retained the highest level of calibration.

Interpretation

Only six models can be directly applied to the Yinzhou regional health care database. The JPHC model developed for the Japanese population and the LCRS model developed based on the China Kadoorie Biobank (CKB) performed better in the Chinese population than other models.

Funding

This work was supported by the National Natural Science Foundation of China (82473728 to Y.W.) and Medical and Health Science and Technology Project of Zhejiang Province, China.