Risk Prediction Models for Lung Cancer: A Systematic Review

Abstract

Many lung cancer risk prediction models have been published but there has been no systematic review or comprehensive assessment of these models to assess how they could be used in screening. We performed a systematic review of lung cancer prediction models and identified 31 articles that related to 25 distinct models, of which 11 considered epidemiological factors only and did not require a clinical input. Another 11 articles focused on models that required a clinical assessment such as a blood test or scan, and 8 articles considered the 2-stage clonal expansion model. More of the epidemiological models had been externally validated than the more recent clinical assessment models. There was varying discrimination, the ability of a model to distinguish between cases and controls, with an area under the curve between 0.57 and 0.879 and calibration, the model's ability to assign an accurate probability to an individual. In our review we found that further validation studies need to be considered; especially for the Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening Trial 2012 Model Version (PLCOM2012) and Hoggart models, which recorded the best overall performance. Future studies will need to focus on prediction rules, such as optimal risk thresholds, for models for selective screening trials. Only 3 validation studies considered prediction rules when validating the models and overall the models were validated using varied tests in distinct populations, which made direct comparisons difficult. To improve this, multiple models need to be tested on the same data set with considerations for sensitivity, specificity, model accuracy, and positive predictive values at the optimal risk thresholds.

Keywords: Clinical utility; Prediction models; Prediction models' design; Prediction models' discrimination; Validation results.