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

Area of Science:

• Critical Care Medicine
• Machine Learning in Healthcare
• Thrombosis Research

Background:

• Deep vein thrombosis (DVT) is a significant risk for critically ill patients in the Intensive Care Unit (ICU).
• Existing risk assessment tools often lack accuracy for predicting DVT in the ICU setting.
• There is a need for improved, interpretable models for DVT risk stratification in critically ill populations.

Purpose of the Study:

• To develop and externally validate an interpretable machine learning (ML) model for predicting DVT risk in ICU patients.
• To identify key clinical predictors of ICU-acquired DVT.
• To compare the performance of ML models against traditional risk assessment tools.

Main Methods:

• A multicenter retrospective study involving 2000 ICU patients.
• Development and validation of eight ML algorithms, including Random Forest (RF).
• Utilized LASSO regression, Boruta algorithm for feature selection, and SHAP for model interpretability.
• Evaluated models using Area Under the Curve (AUC), calibration plots, and Decision Curve Analysis (DCA).

Main Results:

• The RF model demonstrated superior performance with AUCs of 0.869 (training), 0.850 (internal validation), and 0.831 (external validation).
• Key predictors identified by SHAP analysis include immobilization duration, D-dimer levels, femoral vein catheterization, APACHE II score, malignancy, and age.
• The RF model effectively captured non-linear relationships, showing exponential risk increases with prolonged immobilization and high D-dimer.
• DCA indicated a higher net clinical benefit for the RF model compared to standard strategies.

Conclusions:

• An interpretable RF model integrating six accessible clinical variables provides robust DVT risk stratification in ICU patients, surpassing traditional scores.
• The developed model offers a promising tool for personalized thromboprophylaxis strategies in critical care.
• A web-based calculator is under development to facilitate clinical implementation and improve early DVT prevention.