Head-to-head comparisons of risk discrimination by questionnaire-based lung cancer risk prediction models: a systematic review and meta-analysis.

Abstract

Background: While different lung cancer risk prediction models have been established as essential tools to identify high-risk participants for lung cancer screening programs, evaluations of their risk discriminatory performances have reported heterogenous findings in different research cohorts. We therefore aimed to summarise results of head-to-head comparisons of the predictive performance of various lung cancer risk models performed within the same study population.

Methods: In this systematic review and meta-analysis, we performed a systematic search of PubMed and Web of Science databases for primary studies published from inception to Oct 16, 2024. Articles comparing the performance of questionnaire-based lung cancer risk models in an independent, external validation cohort of participants with previous or current smoking exposure were included. The main reasons for exclusion of studies were if only one model was assessed in the external population or risk discrimination was not evaluated. Random-effects meta-analyses were conducted to synthesize differences in the area under the curve (AUC) of two models compared in multiple populations. To assess the risk of bias, PROBAST (the Prediction model Risk of Bias Assessment Tool) was used. The study was registered with PROSPERO, CRD42023427911.

Findings: The systematic search yielded 5568 records. In total, 15 eligible studies were included in the meta-analysis, comprising 4,134,648 individuals with previous or current smoking exposure, of whom 45,448 (1.10%) developed LC within 5-7 years. Among the nine models that were compared, AUC differences reached up to 0.050 between two models. The Lung Cancer Risk Assessment Tool (LCRAT), Bach model and PLCOm2012 model consistently had a higher AUC when compared to any other model, with AUC differences ranging between 0.018 (95% CI 0.011, 0.026) and 0.044 (95% CI 0.038, 0.049). The risk of bias and applicability concerns were deemed low in eight, and high in seven of the included studies. Results excluding studies with high risk of bias were mostly consistent. Among eight of the 24 model pairs that were compared, there was notable between-study heterogeneity (I² ≥50%).

Interpretation: Our systematic review and meta-analyses of head-to-head comparisons disclose major differences in predictive performance of widely used lung cancer risk models. Although our review is limited to the availability of head-to-head comparisons, evidence from current cohort-based model comparisons indicates that the LCRAT, Bach and PLCOm2012 consistently outperformed alternative questionnaire-based risk prediction tools.

Funding: Funded by the European Union.