Lung cancer risk assessment by prediction model: a global perspective.

Abstract

Numerous lung cancer risk prediction models have been developed and validated worldwide. It is imperative to offer a comprehensive overview and comparative analysis of their performances. We conducted an extensive literature search to identify studies developing and/or validating lung cancer risk prediction models. Then we summarised and compared the external performance of these models, focusing on discriminative accuracy (C-index) and calibration performance (E:O ratio). After an initial screening of 10 210 articles, 35 studies on 21 distinct prediction models were identified, which used 42 different types of predictors spanning seven categories. Notable performance variations were observed in external validations. In North American cohorts, the C-index ranged from 0.60 to 0.87, with E:O ratios from 0.62 to 3.70. Among the European cohorts, the Trøndelag health study HUNT and CanPredict exhibited C-indices surpassing 0.870. Conversely, the Bach, lung cancer risk assessment tool (LCRAT), prostate, lung, colorectal and ovarian cancer screening (PLCO)_m2012 and PLCO_all2014 performed poorly in electronic health records of the Qresearch database subgroup, with C-indices falling below 0.60. PLCO_m2012 reached the best E:O ratio of 1.00 (95% CI: 0.93 to 1.08) in the UK Biobank subgroup. In Asian cohorts, the C-index ranged from 0.54 to 0.87. Only three models, Korean Men, LCRAT and Liverpool lung project incidence risk model (LLPi), achieved a C-index exceeding 0.80. LCRAT demonstrated the best calibration, while Hoggart performed the worst. Performance of lung cancer risk prediction models, despite being well developed and validated, varies in diverse populations. Significant regional imbalance persists in the development of these models. Rigorous external validation or recalibration study in the target population is crucial in accordance with the guidance prior to model implementation. CRD42022324602.