Improving lung cancer screening: the role and challenges of risk prediction models

Abstract

Targeted lung cancer screening saves lives by shifting diagnosis to earlier stages, when curative treatment is more likely to succeed. To improve the efficiency of screening and minimise associated harms, tools have been developed to identify individuals at elevated risk. Most major clinical trials defined eligibility using categorical criteria, typically based on age and cumulative smoking exposure—the method adopted in US guidelines since 2013.1 More recently, personalised risk prediction models have been developed to estimate an individual’s probability of developing lung cancer, with the aim of further enhancing screening performance. These models are now in clinical use in countries including England, Canada and Australia.2 The systematic review by Zhang et al offers a comprehensive analysis of performance measures for 21 risk models from 35 studies, with an appropriate emphasis on external validation to mitigate bias.3 Crucially, the review highlights substantial variation in how model performance is reported. In many cases, key metrics such as model calibration, often presented as the expected-to-observed ratio (E:O ratio), are missing altogether. Additional assessments such as calibration slope and decision curve analysis are reported even less frequently, which likely explains their exclusion from the review. The clinical utility …