A more selective risk-stratified, machine learning based, lung cancer thromboprophylaxis protocol following vats
Jing Wang1,2, Di Wang3, Songping Cui2
1Mass General Cancer Center, Massachusetts General Hospital, Harvard Medical School, Boston, MA.
Background:
Venous thromboembolism (VTE) represents a potentially fatal but preventable postoperative complication. We sought to establish and validate an explainable prediction model based on the machine learning (ML) approach for VTE, and assess its prognostic implications in thoracic oncology patients undergoing VATS Segmentectomy.
Materials And Methods:
We prospectively developed and validated a predictive model for postoperative VTE following VATS segmentectomy. Patients were sequentially enrolled into training (n = 557, Apr.2017-Jan.2021) and validation cohorts (n = 239, Feb.2021-Oct.2022). 49 clinicopathological variables, including the novel biomarker von Willebrand factor A2 (vWF-A2), were evaluated. 11 ML algorithms were compared based on several evaluation indexes including AUC. SHapley Additive exPlanations (SHAP) analysis was utilized for feature ranking and interpretability. The final model was benchmarked against the traditional Caprini score, and the prognostic impact of postoperative VTE on long-term survival was further assessed.
Results:
In this prospective study, eXtreme gradient boosting (XGBoost) demonstrated superior discriminative performance among 11 evaluated ML-models. After feature reduction based on ranked feature importance, a final interpretable XGBoost model comprising 11 variables was established. This model accurately predicted postoperative VTE in both training (AUC = 0.903) and validation (AUC = 0.856) cohorts, significantly outperforming the conventional Caprini score RAM. Additionally, comparison of oncologic outcomes revealed no significant difference in overall survival (P = 0.068), whereas disease-free survival was significantly shorter in patients experiencing postoperative VTE (P = 0.017).
Conclusion:
Our explainable risk-stratification ML model not only accurately predicts the risk of VTE following VATS segmentectomy in early-stage NSCLC patients, but also exhibits substantial clinical relevance to adverse prognostic outcomes in this patient cohort.
