Machine Learning-Based Early Prediction of Lower Extremity Deep Vein Thrombosis in the ICU: A Multicenter Study.

Abstract

Purpose: Deep vein thrombosis (DVT) is a prevalent and life-threatening complication in the Intensive Care Unit (ICU). Traditional risk tools often lack specificity for ICU acquired DVT. This study aimed to develop and externally validate an interpretable machine learning (ML) model for accurate DVT risk prediction in critically ill patients.

Methods: This multicenter retrospective study included 2000 patients from two centers, partitioned into training, internal testing, and external validation sets. LASSO regression and the Boruta algorithm identified robust predictors. Eight ML algorithms were trained and evaluated using the Area Under the Curve (AUC), calibration plots, and Decision Curve Analysis (DCA). The SHapley Additive exPlanations (SHAP) framework was utilized for model interpretability.

Results: The Random Forest (RF) model outperformed other ML algorithms and traditional tools (e.g. Caprini score). It achieved AUCs of 0.869 (95% CI: 0.822-0.916) in training, 0.850 (95% CI: 0.801-0.900) in internal testing, and 0.831 (95% CI: 0.771-0.888) in external validation. SHAP analysis identified six dominant predictors: immobilization duration, D-dimer, femoral vein catheterization, APACHE II score, malignancy, and age. The RF model successfully captured non-linear interactions, particularly exponential risk increases from prolonged immobilization and elevated D-dimer. DCA demonstrated a higher net clinical benefit than default strategies.

Conclusion: Integrating six readily available clinical variables, the RF model offers a robust and interpretable tool for DVT risk stratification, outperforming traditional scores. To facilitate real-world clinical application, an accessible web-based calculator is being developed to guide early, personalized thromboprophylaxis in the ICU.