Predicting the future risk of lung cancer: Development and validation of QCancer2 (10-year risk) lung model and evaluating the model performance of nine prediction models

Abstract

Objectives: To develop and validate the QCancer2 (10-year risk) lung model for estimation of future risk of lung cancer and to compare the model performance against other prediction models for lung cancer screening Design: open cohort study using linked electronic health records (EHRs) from the QResearch database (1 January 2005 - 31 March 2020) Setting: English primary care Participants: 12.99 million patients aged 25-84 years were in the derivation cohort to develop the models and 4.14 million patients were in the validation cohort. All patients were free of lung cancer at baseline. Main Outcome Measure(s): Incident lung cancer cases Methods: There were two stages in this study. First, Cox proportional hazards models were used in the derivation cohort to update the QCancer (10-year risk) lung model in men and women for a 10-year predictive horizon, including two new predictors (pneumonia and venous thromboembolism) and more recent data. Discrimination measures (Harrell's C, D statistic, and RD2) and calibration plots were used to evaluate model performance in the validation cohort by sex. Secondly, seven prediction models for lung cancer screening (LLPv2, LLPv3, LCRAT, PLCOM2012, PLCOM2014, Pittsburgh, and Bach) were selected to compare the model performance with the QCancer2 (10-year risk) lung model in two subgroups: (1) smokers and non-smokers aged 40-84 years and (2) ever-smokers aged 55-74 years. Result(s): 73,380 incident lung cancer cases were identified in the derivation cohort and 22,838 in the validation cohort during follow-up. The updated models explained 65% of the variation in time to diagnosis of lung cancer (RD2) in both sexes. Harrell's C statistics were close to 0.9 (indicating excellent discrimination), and the D statistics were around 2.8. Compared with the original models, the discrimination measures in the updated models improved slightly in both sexes. Compared with other prediction models, the QCancer2 (10-year risk) lung model had the best model performance in discrimination, calibration, and net benefit across three predictive horizons (5, 6, and 10 years) in the two subgroups. Conclusion(s): Developed and validated using large-scale EHRs, the QCancer2 (10-year risk) lung model can estimate the risk of an individual patient aged 25-84 years for up to 10 years. It has the best model performance among other prediction models. It has potential utility for risk stratification of the English primary care population and selection of eligible people at high risk for the targeted lung health check programme or lung cancer screening.Copyright The copyright holder for this preprint is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. All rights reserved. No reuse allowed without permission.