Development and validation of a machine-learning-based triage model incorporating systemic inflammatory, nutritional, coagulation, and tumor marker parameters for identifying baseline distant metastasis in lung cancer.

Purpose: This study aimed to develop and validate a machine learning-based model to identify baseline distant metastasis in lung cancer by integrating multidimensional pretreatment indicators-including systemic inflammatory, tumor, coagulation, and nutritional parameters. Methods: This retrospective case-control study enrolled 408 treatment naïve patients with newly diagnosed lung cancer at Guangzhou First People's Hospital between November 2023 and October 2024, including 202 cases with and 206 without distant metastasis. A total of 18 peripheral blood biomarkers-encompassing systemic inflammation, tumor burden, coagulation, and nutritional status-along with six demographic and clinical characteristics, were collected as candidate predictors. After initial univariate logistic regression screening, nested 5-fold cross-validation was implemented within the training set, within each fold of which the three complementary feature selection algorithms (LASSO, RFE, and Boruta) were independently applied in combination to derive a robust set of predictive features. Based on the selected features, nine machine learning classifiers were then developed within each fold, and algorithm comparison as well as optimal classifier selection were performed using the mean AUC across the five folds. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, positive predictive value, negative predictive value, F1 score, and 95% confidence interval. The optimal model was selected based on overall performance; its calibration was assessed using calibration plots and the Hosmer-Lemeshow goodness of fit test, while clinical net benefit was quantified through decision curve and clinical impact curves. To facilitate bedside risk stratification, a nomogram was constructed based on the final model to enable individualized probability estimation. Feature importance and model interpretability were further examined using SHAP values. Results: The AdaBoost algorithm demonstrated the best performance for identifying baseline distant metastasis in lung cancer, achieving a test AUC of 0.840 (95% CI: 0.757-0.912), and was selected as the final model. A rigorous nested 5-fold cross-validation provided an unbiased AUC estimate of 0.82, which aligns closely with the test AUC of 0.84 and further confirms the model's stability. Learning curves, decision curve analysis, and calibration curves collectively confirmed its favorable generalization, clinical net benefit, and calibration accuracy, with positive net benefit across a wide range of threshold probabilities. SHAP analysis identified CEA, SII, smoking history, CYFRA21-1, D-dimer, fibrinogen, AGR, histological type, SIRI, and NSE as the top 10 contributors to the model's risk stratification-rankings that closely aligned with those from the decision tree model-underscoring their biological and clinical relevance. Conclusion: This study developed and internally validated an AdaBoost-based model for pretreatment identification of baseline distant metastasis in lung cancer. As a low-cost, non-invasive tool integrating routine laboratory parameters, it shows preliminary promise for early risk stratification, although external validation is required before clinical implementation.