Machine Learning for Classification in Lung Cancer Using Routine Clinical and Laboratory Data

Abstract Background Accurate pathological classification of lung cancer is essential for informing treatment strategies. However, invasive biopsy procedures are not feasible for high-risk patients or those with inaccessible lesions. This study aimed to develop a machine learning model utilizing routine clinical and laboratory data for classification of non-invasive lung cancer. Methods Data from patients admitted to Sichuan Provincial Cancer Hospital were retrospectively analyzed. Key features were determined using LASSO and Boruta algorithms. Four machine learning models, including logistic regression, extreme gradient boosting (XGBoost), categorical boosting (CatBoost), and random forest (RandomForest), were trained and optimized through five-fold cross-validation. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, and F1 score. An online calculator was developed using R Shiny for clinical deployment. Results A total of 1122 patients with lung cancer were included and randomly assigned to the training and test sets. In the training set, 16 features were incorporated into the models. The RandomForest model demonstrated superior performance compared with the other models, achieving an AUC of 0.999, an accuracy of 0.984, and an F1 score of 1.000. Notably, sex and tumor markers were identified as significant predictors. In the test set, the RandomForest model attained a micro-averaged AUC of 0.969 and macro-averaged AUC of 0.940. Sensitivity and specificity varied from 0.667 to 0.995 across subtypes. A web-based tool was implemented to facilitate real-time clinical application ( https://nkuwangkai.shinyapps.io/lung-cancer-v1/ ). Conclusion This study presented a robust, non-invasive machine learning model for lung cancer subtype classification, addressing critical gaps in clinical practice for biopsy-ineligible patients. A web-based calculator was developed to facilitate clinical application. Nonetheless, future multicenter validation is warranted to expand the generalizability of this model and promote adoption in diverse healthcare settings.