Predicting the risk of lung cancer using machine learning: A large study based on UK Biobank

Abstract

In response to the high incidence and poor prognosis of lung cancer, this study tends to develop a generalizable lung-cancer prediction model by using machine learning to define high-risk groups and realize the early identification and prevention of lung cancer. We included 467,888 participants from UK Biobank, using lung cancer incidence as an outcome variable, including 49 previously known high-risk factors and less studied or unstudied predictors. We developed multivariate prediction models using multiple machine learning models, namely logistic regression, naïve Bayes, random forest, and extreme gradient boosting models. The performance of the models was evaluated by calculating the areas under their receiver operating characteristic curves, Brier loss, log loss, precision, recall, and F1 scores. The Shapley additive explanations interpreter was used to visualize the models. Three were ultimately 4299 cases of lung cancer that were diagnosed in our sample. The model containing all the predictors had good predictive power, and the extreme gradient boosting model had the best performance with an area under curve of 0.998. New important predictive factors for lung cancer were also identified, namely hip circumference, waist circumference, number of cigarettes previously smoked daily, neuroticism score, age, and forced expiratory volume in 1 second. The predictive model established by incorporating novel predictive factors can be of value in the early identification of lung cancer. It may be helpful in stratifying individuals and selecting those at higher risk for inclusion in screening programs.