Development and validation of a machine learning model for predicting preoperative deep vein thrombosis in elderly hip fracture patients

Figure 1. Flowchart of data screening and analysis.

Figure 2. Comparative performance evaluation of machine learning models. Performance evaluation of five machine learning models—decision tree (DT), light gradient boosting machine (LGBM), logistic regression (LR), support vector machine (SVM), and extreme gradient boosting (XGBoost)—in the training and validation datasets. (A,B) ROC curves with AUC values, showing XGBoost’s superior discriminative performance. (C,D) Calibration curves indicating a strong fit and robust generalization for XGBoost. (E,F) DCA curves demonstrating XGBoost’s net clinical benefit. ROC, receiver operating characteristic; DCA, decision curve analysis; AUC, area under the curve; DT, decision tree; LR, logistic regression; SVM, support vector machine; LGBM, light gradient boosting machine; XGBoost, extreme gradient boosting. Panels (A,C,E) represent results from the training dataset, whereas Panels (B,D,F) correspond to results from the validation dataset.

Figure 3. (A) SHAP feature importance plot. The SHAP summary plot ranks feature importance in the XGBoost model based on mean SHAP values. Feature values are indicated by a color gradient, with yellow representing high values and purple representing low values. (B) SHAP dependence plot. The SHAP dependence plots illustrate the marginal effects of five key features on the model’s predictions, highlighting the relationships between deep vein thrombosis (DVT) and its predictive factors. (C) SHAP force plot for a true positive case. The SHAP force plot for a true positive patient shows a SHAP value of −0.686, exceeding the model’s baseline (E[f(x)] = −1.22). Positive contributions from Hemoglobin, Injury-to-admission time, D-dimer, and (in yellow) increased the predicted likelihood of DVT, while Albumin and APTT (in red) exerted a negative influence, reducing the predicted risk. (D) SHAP force plot for a true negative case. In contrast, the SHAP force plot for a true negative patient yielded a SHAP value of −2.59, below the baseline. Negative contributions from D-dimer and Albumin (in red) reduced the predicted probability of DVT, whereas minor positive contributions from Hemoglobin, APTT, and…

Figure 4. An online platform has been developed to predict the risk of preoperative deep vein thrombosis (DVT) in elderly patients with hip fractures. The web-based tool, accessible via https://3123213wxk.shinyapps.io/1111/, enables clinicians to perform real-time risk assessments by entering individual patient data. Through an interactive visual interface, the application estimates the likelihood of DVT occurrence based on key clinical characteristics.