OA2-AM24-MN-08 - Machine Learning Versus Traditional Regression Analytics to Study the Association between Peri-Operative Red Blood Cell Transfusions and Venous Thromboembolism

We used ­machine learning analytics to explore the relationship between perioperative RBC transfusions and post-operative venous thromboembolism (VTE); deep venous thrombosis and/or pulmonary embolism) within 30 days of surgery using a large prospective validated surgical registry – American College of Surgeons’ National Surgical Quality Improvement Project (ACS-NSQIP).  This database collates data from more than 525 teaching/non-teaching hospitals in North America. Greater than 1 million patient records from 2018 and 2019 were used as model development and validation cohorts respectively. Clinically confirmed VTE cases warranting a therapeutic intervention were included in the analysis.

A segmented modeling approach was developed in two stages: 1) patient segmentation for homogeneous characteristics using unsupervised clustering: K-means clustering was applied for data-driven patient segmentation, optimizing within-cluster similarity across various cluster counts for the optimal number of clusters; 2) Clustering-based prediction models: For each cluster, predictive models were developed and compared using lasso logistic regression and eXtreme Gradient Boosting (XGBoost). In contrast, a traditional, unified approach was developed by applying lasso logistic regression and XGBoost across all patients, bypassing patient clustering analysis.

Results/Findings: In the development cohort, 47,091 (4.6%) received at least one RBC transfusion perioperatively (PreOp 0.4%; Intra/Postop 3.8% and both Pre/Intra/Postop 0.4%). Postop new/progressive VTE was identified in 7,749 (0.8%). Among the 67 predictors used for model development, RBC transfusions emerged as a top predictor for postoperative VTE prediction using unified or segmented approach. The unified approach showed perioperative RBC transfusions were associated with higher risk of VTE (odds ratio (OR)(95%CI)=1.81 (1.69-1.94)) overall in both the development and validation cohorts. Patient segmentation revealed cluster-specific ORs ranging from 1.01 (0.72 - 1.40) to 1.51 (1.39 - 1.63) (Table 1). In the unified approach, XGBoost outperformed lasso logistic regression in predictive accuracy, while both models performed similarly in the segmented approach.

Conclusions: These data suggest association of perioperative RBC transfusions with the development of new/progressive post-operative VTE.  The traditional regression analysis based unified approach oversimplified heterogeneous data and failed to capture differential effects. The innovative segmented approach using machine learning analytics tailored the models to distinct patient groups within a large dataset, thus identifying variable effect sizes for this association which enhanced both predictive accuracy and interpretability.