Risk prediction of non-small cell lung cancer in patients with pulmonary nodules: a single-center cohort study based on six machine learning algorithms.

Lung cancer is a major global public health concern. Non-small cell lung cancer (NSCLC), the most common subtype, has drawn increasing attention. To achieve early and accurate identification of NSCLC in patients with pulmonary nodules, reliable clinical prediction models are required. This study aimed to establish a machine learning model that integrates imaging and clinical features to predict the risk of NSCLC in this population. Data from 963 patients with pulmonary nodules, treated at our hospital between January 2020 and December 2024, were retrospectively collected. Patients were randomly divided into training and testing cohorts at a 7:3 ratio. Demographic, clinical, and imaging variables were included. Two algorithms were applied for variable selection, and selected variables were incorporated into the final models. Six machine learning algorithms were trained with tenfold cross-validation. The mean area under the receiver operating characteristic curve (AUC) was used to identify the best model. The top-performing model was further evaluated, and interpretability was analyzed using SHapley Additive exPlanations (SHAP). A web-based calculator was then developed for clinical application. The XGBoost model showed the best performance among all machine learning methods in the training, testing, and validation cohorts. It achieved an AUC of 0.943 in the training cohort and 0.936 in the testing cohort. The most influential predictors were ground-glass opacity (GGO), nodule density, hypertension, plasma fibrinogen, blood urea nitrogen (BUN), and nodule size. The final model was implemented as a web-based clinical calculator ( https://qq568325999.shinyapps.io/dynnomapp/ ) to facilitate clinical use. The XGBoost model developed in this study showed good discriminative ability for predicting the risk of NSCLC in patients with pulmonary nodules. These findings suggest that the model may have potential clinical utility. It may serve as an adjunctive tool for early risk stratification in this patient population. The use of SHAP analysis further enhanced the interpretability of the model. In addition, the online calculator derived from this model may facilitate its use in clinical research and future clinical translation. However, this study was a single-center retrospective analysis and included only internal validation. Therefore, further validation of the model is needed in multicenter external cohorts and prospective studies.