The venous thromboembolism problem Venous thromboembolism (VTE) - encompassing deep vein thrombosis and pulmonary embolism - is a major complication after lung cancer surgery. NSCLC patients have an annual VTE incidence of up to 20%, and surgery increases VTE and pulmonary embolism-related death risk by 2- and 4-fold respectively.
Limitations of current risk scores Existing VTE risk models including Caprini and Khorana were developed in Western Caucasian populations and perform poorly in Chinese patients (AUC as low as 0.474 for the modified Caprini score in thoracic surgery). None have incorporated lung cancer driver mutations as risk factors.
EGFR mutation and VTE EGFR mutations occur in 44-63% of Asian NSCLC patients and are associated with altered coagulation biology, but the relationship between EGFR mutation status and postoperative VTE risk has not been well established.
Study approach Published in PLOS ONE (2025), this retrospective study of 362 NSCLC surgery patients develops and validates a stacking machine learning model using EGFR mutation status alongside clinical biomarkers to predict failure of standard thromboprophylaxis.
Study population The retrospective cohort included 362 NSCLC adenocarcinoma patients who underwent lobectomy and systemic lymph node resection at Xingtai People's Hospital between January 2018 and November 2022. All received 28 days of nadroparin thromboprophylaxis post-surgery.
VTE incidence and definition VTE occurred in 58 patients (16.0%). Deep vein thrombosis was screened by weekly lower-extremity ultrasound for 4 weeks post-surgery, with CT pulmonary angiography for suspected pulmonary embolism.
Variables collected Data included age, sex, smoking status, clinical stage, EGFR mutation status (wild-type vs. exon 19, 20, 21 mutations), complete blood counts, coagulation parameters (D-dimer, fibrinogen, APTT), albumin, fasting glucose, and Caprini scores.
Statistical filtering Multivariable logistic regression identified independent VTE predictors. Age, platelets, D-dimers, albumin, smoking history, and EGFR exon 21 mutation status were selected as final model features with p-values under 0.05.
Nine base models trained Nine machine learning classifiers were trained on the 70% training set: XGBoost, LightGBM, Random Forest, AdaBoost, Gaussian Naive Bayes, Logistic Regression, MLP Neural Network, Support Vector Classifier, and K-Nearest Neighbors. 5-fold cross-validation optimized hyperparameters for each.
Stacking architecture The three highest-performing base models - LightGBM Classifier, Random Forest Classifier, and Gaussian Naive Bayes - were selected as the first-layer learners. Their out-of-fold predictions became features for a second-layer logistic regression meta-learner.
Why stacking works Stacking combines models with different inductive biases, allowing the meta-learner to learn which base model's predictions are most reliable under different feature profiles. This typically outperforms any individual model or simple averaging ensemble.
Evaluation approach The stacking model was evaluated on the held-out 30% test set and a separate validation set, with performance measured by ROC-AUC, accuracy, sensitivity, and specificity. Decision curve analysis assessed clinical net benefit.
Stacking model performance The stacking model achieved ROC-AUC of 0.984 in training and 0.979 in the test set, with accuracy of 0.954, sensitivity of 1.000, and specificity of 0.887. In the external validation set, AUC was 0.983 with sensitivity of 0.978.
EGFR exon 21 significance EGFR exon 21 mutation (L858R) was identified as an independent predictor of VTE risk after surgery. This is clinically important because exon 21 mutations are the most common EGFR alteration in Asian patients and are not currently factored into VTE risk scores.
Decision curve analysis Decision curve analysis showed the stacking model provided high net clinical benefit across a wide range of threshold probabilities for VTE risk intervention, validating its practical utility for guiding individualized prophylaxis decisions.
Superior to Caprini score The machine learning model significantly outperformed the modified Caprini score in the study population, confirming that Asian NSCLC patients are not well-served by existing Western-derived risk tools.
Precision thromboprophylaxis The model could identify which NSCLC surgical patients are at highest risk for thromboprophylaxis failure, enabling extended anticoagulation duration, dose escalation, or more intensive VTE surveillance for high-risk individuals.
EGFR testing integration Since EGFR mutation testing is routine in NSCLC patients prior to treatment, EGFR status is already available at the time of surgery. Incorporating this result into VTE risk calculations requires no additional testing.
Asian patient applicability The model was developed specifically in a Chinese NSCLC patient cohort, making it more appropriate for this population than existing Western-derived risk tools - directly addressing an equity gap in current evidence-based medicine.
Routine clinical deployment The six-variable model uses data routinely collected in pre-surgical workup (age, platelet count, D-dimer, albumin, smoking history, EGFR status), making deployment feasible without requiring additional data collection workflows.
Single-center retrospective design Results require external validation in diverse hospitals and healthcare systems before clinical implementation. Performance may differ in populations with different EGFR mutation frequencies and VTE baseline rates.
Limited sample size With 362 patients and 58 VTE events, the dataset is relatively small for a complex multi-class machine learning problem. Larger multi-center cohorts would provide more stable model estimates.
Neoadjuvant therapy exclusion Patients who received pre-surgical neoadjuvant therapy were excluded. VTE risk in this growing patient population - increasingly common with immunotherapy and targeted therapy use - requires separate study.
Prospective validation The retrospective design cannot confirm that using the model to guide VTE prophylaxis decisions would improve clinical outcomes. A prospective interventional trial testing model-guided versus standard prophylaxis is the needed next step.