Clinical risk factors and machine learning for venous thromboembolism in lung cancer: an exploratory single-center...

Venous thromboembolism (VTE) represents a common and severe comorbidity in lung cancer patients. This study aimed to characterize the clinical features and independent risk factors associated with VTE in these individuals, and to evaluate the complementary utility of machine learning approaches in enhancing VTE risk stratification.

This retrospective study included lung cancer patients admitted to the Fourth Affiliated Hospital of Anhui Medical University between January 2018 and April 2025. According to imaging findings, patients were categorized into a VTE group and a control group. Clinical characteristics, laboratory parameters, and treatment-related information were retrospectively collected and analyzed. Univariate and multivariable logistic regression analyses were performed to identify factors independently associated with VTE. Additionally, four machine learning models-Ridge-LR, LASSO, random forest (RF), and extreme gradient boosting (XGBoost)-were developed for exploratory purposes. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, calibration metrics (Brier score, calibration slope and intercept, Hosmer-Lemeshow test), and decision curve analysis (DCA).

A total of 180 patients with lung cancer were included. Elevated D-dimer levels were independently associated with increased VTE risk (OR = 1.084, p = 0.021). Chemotherapy exposure showed an OR of 2.858 (95% CI: 1.081-7.555, p = 0.034) for VTE in the 1-2 cycle subgroup, although no significant association was observed for higher cycle numbers, suggesting potential confounding. Male sex was independently associated with embolic events occurring outside the lower extremities (OR = 10.83, p = 0.034). Among the machine learning models, RF achieved the highest AUC (0.740), but Ridge-LR demonstrated the best calibration (Hosmer-Lemeshow p = 0.345). LASSO showed the highest net benefit on decision curve analysis, while RF and XGBoost exhibited significant miscalibration (p = 0.002 for both). D-dimer (41.3%) and total protein (34.7%) were identified as the most important predictors in the XGBoost model.

Elevated D-dimer levels measured prior to VTE diagnosis during routine clinical monitoring were independently associated with subsequent VTE, supporting its potential utility as a predictive biomarker for risk stratification in patients with lung cancer. The association between chemotherapy exposure and VTE is likely influenced by selection bias and residual confounding rather than representing a true causal effect. Machine learning models did not demonstrate consistent improvement over logistic regression in discrimination or calibration in this small single-center cohort, suggesting limited added predictive value with the current dataset.