A database for lung cancer prediction models - a Danish cohort study of nearly 40,000 patients

Abstract Lung cancer (LC) is the leading cause of cancer related deaths due to late-stage diagnosis resulting in a poor prognosis. Several risk models have been presented to refine LC screening, although mainly based on unrepresentative populations and limited clinical parameters. In order to create accurate risk models targeting the relevant population, systematic data collection and data preparation should always consider the clinical application and relevant population at risk. Based on a combination of multiple data sources this study aims to describe differences between LC patients and non-LC patients in a population at risk. This will serve as the foundation for future analysis in the creation of prediction models. All patients were included who underwent examinations on suspicion of LC in the Region of Southern Denmark during 2009-2018. A large variety of clinical and laboratory data were collected and distributions between the LC and non-LC cohorts were compared using standard statistical methods. The number of patients examined in the LC fast-track clinics increased substantially over the study period from 2,806 in 2009 to 4,740 in 2018. The proportion of LC patients decreased from 30% in 2009 to 25% in 2018. More patients were diagnosed in stage I-II, from 18% in 2009 to 31% in 2018. The stereotypical LC patient is likely to be an around 70-year old female, smoker with a history of cardiovascular disease, chronic obstructive pulmonary disorder and blood analysis indicating inflammation, hyponatremia and hypoalbuminemia. Considering missing results and applying relevant filters resulted in a cohort of 9,940 patients with complete results of all datasets. The patients of the study cohort share many baseline risk factors and are all considered to be at risk of LC. However, several differences between LC patients and non-LC patients were found. Using methods based on artificial intelligence, we will exploit these differences in order to create risk estimates of specific patients. A high-performing risk model providing decision support to the general practitioner would facilitate earlier referral of potential LC-patients.