Development and internal validation of a machine learning-based clinical prediction model for liver metastasis in stage III–IV non-small cell lung cancer

Abstract

Liver metastasis is among the major determinants of prognosis in non-small cell lung cancer (NSCLC), and early identification of high-risk patients has direct implications for clinical management. We retrospectively enrolled 849 patients with stage III–IV NSCLC and split them 8:2 into a training set ( n = 679) and an internal validation set ( n = 170). Feature selection followed a Venn intersection approach combining Least Absolute Shrinkage and Selection Operator (LASSO) regression, random forest (RF), and decision tree (DT) algorithms, which identified six shared predictors: lactate dehydrogenase (LDH), multi-organ metastasis, hemoglobin (HGB), thrombin time (TT), platelet count (PLT), and alkaline phosphatase (ALP). Ten machine learning (ML) classifiers were trained on these variables, and two with complementary strengths were assembled into a Stacking ensemble. On the internal validation set, the Stacking model outperformed all candidate models (AUC = 0.852, 95% CI: 0.769–0.921; AUPRC = 0.565; sensitivity = 76.0%; NPV = 95.2%). Calibration curves showed well-calibrated probability estimates, and decision curve analysis demonstrated net clinical benefit. SHAP analysis identified elevated LDH (roughly ≥ 250 U/L) and multi-organ metastasis as the two strongest contributors to individual predictions, while lower HGB, shortened TT, and abnormal PLT or ALP each added smaller but independent predictive weight. These findings demonstrate that a Stacking ML approach can predict liver metastasis risk in stage III–IV NSCLC with strong discrimination and reliable calibration. An online calculator has been deployed for point-of-care use; external validation in prospective cohorts is needed before broader clinical application.