Predicting deep vein thrombosis using machine learning and blood routine analysis

Figures

• 

  Figure 1. Flowchart of the ML development model for predicting DVT.

• 

  Figure 2. Analysis of the distribution of hospitalized patients with DVT.

• 

  Figure 3. Diagram of the Screening and Optimization Process for SHAP Feature Impact Distribution. (A) SHAP Summary Plot of the First-Round Feature Selection. Color Coding: The color gradient from blue to red represents feature values, where blue indicates a low feature value and red indicates a high feature value. X-axis: Denotes the SHAP value, which reflects the impact of each feature on the output of DVT risk prediction model. Features labeled in the plot were excluded in the first round due to their SHAP value range being less than 0.2. (B) SHAP Summary Plot of the Second-Round Feature Selection. Low-contribution features highlighted in the plot were removed in the second round to refine the feature set. (C) SHAP Summary Plot of the Final-Round Feature Selection. Clinical Annotations for Each Feature: age: Increased age is associated with higher thrombotic susceptibility; RBC: RBC levels are linked to blood viscosity, a known risk factor for DVT; MCV: As an erythrocyte parameter, it is associated with oxygen-carrying capacity, which in turn influences thrombotic risk; PDW: Reflects platelet heterogeneity; abnormal PDW may alter thrombogenic potential; LYMPH%: Associated with…

• 

  Figure 4. Correlation analysis between features.

• 

  Figure 5. ROC curves of six machine-learning models for DVT prediction. The diagonal gray dashed line indicates the reference of a random classifier (AUC = 0.5). AUC with 95% confidence intervals are listed in the legend. RF achieves the highest discriminative performance (AUC = 0.856).

• 

  Figure 6. Calibration curve of the random forest model on the validation set with a Brier score of 0.158.

• 

  Figure 7. Determination of optimal threshold for RF model. A red arrow indicates the optimal threshold point on the ROC curve. The legend shows that the area under the curve (AUC) of the RF model is 0.856, with a 95% confidence interval of 0.810–0.901. At the optimal threshold marked in the figure, the model achieves a sensitivity of 77.97% and a specificity of 78.15%.

• 

  Figure 8. Illustrates the decision curve of the random forest model on the validation set. Within the threshold range of 0.03–0.75, the net benefit achieved by using the model to guide decision-making is consistently superior to that of the two naive strategies, namely predicting all individuals as negative (with a net benefit of 0) and predicting all individuals as positive (with the net benefit corresponding to the “treat all” curve).