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

Figure 1. Feature selection and correlation analysis. (A) LASSO coefficient profiles of candidate variables. (B) Tuning parameter (λ) selection in the LASSO model using 10-fold cross-validation. Vertical lines indicate the minimum error (left) and 1-standard error (right). (C) Feature importance ranking by Boruta algorithm. Green boxplots denote confirmed important features relative to shadow features. (D) Heatmap of Spearman correlation coefficients among the final selected features.

Figure 2. Comparison of different machine learning models in the training cohort. (A) ROC curves with corresponding AUC values and 95% confidence intervals. (B) Heatmap of P-values derived from DeLong’s test for pairwise comparisons of AUCs between models. (C) Bar plot summarizing the AUC scores for each classifier. (D) Comprehensive evaluation heatmap displaying multiple performance metrics (Sensitivity, Specificity, F1-score, etc.) across all models.

Figure 3. Performance evaluation of the Random Forest (RF) model in the Test Cohort. (A) ROC curve illustrating the discrimination performance with an AUC of 0.875. (B) Calibration plot comparing predicted probabilities against observed actual probabilities. The red line represents perfect calibration. (C) Decision curve analysis (DCA) demonstrating the clinical net benefit of the model across different threshold probabilities. (D) ROC curves from 5-fold cross-validation, indicating the model’s stability and robustness.

Figure 4. External validation of the Random Forest (RF) model in the Validation Cohort. (A) ROC curve of the RF model applied to the external validation cohort. (B) Calibration curve showing the agreement between predicted risks and actual outcomes in the validation cohort. (C) Decision Curve Analysis (DCA) demonstrating the clinical utility of the model in the validation cohort. (D) ROC curves showing the generalization performance of the model (with 5-fold validation) in the validation cohort.

Figure 5. SHAP analysis for the interpretability of the optimal Random Forest model. (A) SHAP feature importance plot ranking the global impact of variables on the model’s predictions. (B) SHAP waterfall plot for a representative high-risk patient (f(x)=0.0893), driven by prolonged immobilization and elevated D-dimer. (C) SHAP summary beeswarm plot displaying the distribution and direction of feature effects across the entire cohort. (D) SHAP waterfall plot for a representative low-risk patient (f(x)=−1.13). (E) SHAP waterfall plot for a representative moderate-risk patient (f(x)=0.0405), illustrating the counterbalance between risk-increasing and risk-decreasing factors.

Figure 6. Performance comparison between the optimal Random Forest (RF) model and traditional clinical scoring systems. (A) Receiver operating characteristic (ROC) curves demonstrating the discriminative ability of the models. The RF model achieved the highest area under the curve (AUC = 0.841), outperforming both the Caprini score (AUC = 0.715) and Wells score (AUC = 0.647). (B) Decision curve analysis (DCA) evaluating clinical utility. The RF model provides superior net clinical benefit across the majority of threshold probabilities compared to the two traditional scores and default strategies (treat-all or treat-none).