Machine-learning-derived laboratory signatures associated with postoperative day-3 ultrasound-detected lower-extremity deep vein thrombosis after fractures around the knee

Objective This study aimed to construct risk-discrimination models for postoperative-day-3 ultrasound-detected lower-extremity deep vein thrombosis (DVT) status in patients with fractures around the knee (FAK) using five machine-learning algorithms and compare model performance across five machine-learning algorithms. Methods Electronic medical records of 1,071 patients who underwent surgical treatment for FAK at Suining Central Hospital from January 2023 to August 2026 were retrospectively reviewed. Patients were randomly divided into a training set ( n = 749) and a test set ( n = 322) at a 7:3 ratio. Feature selection occurred exclusively within the training dataset. Predictive variables were screened using least absolute shrinkage and selection operator (LASSO) regression, the Boruta algorithm, and recursive feature elimination-extreme gradient boosting (RFE-XGBoost). Five predictive models were developed, including logistic regression, k-nearest neighbor (KNN), random forest (RF), light gradient-boosting machine (LightGBM), and extreme gradient boosting (XGBoost). Model performance was evaluated comprehensively via the area under the receiver-operating-characteristic curve (AUC), accuracy, sensitivity, specificity, precision, and F1-score. Receiver-operating-characteristic (ROC) curves, calibration curves, and decision-curve analysis (DCA) were utilized to identify the selected model. Finally, Shapley additive explanations (SHAP) were employed to interpret the global and local prediction outputs of the selected model. Results Seventeen overlapping features were identified via feature-selection procedures for final modelling, among which postoperative-day-1 and postoperative-day-3 laboratory biomarkers contributed most substantially to model output. XGBoost was chosen for subsequent SHAP interpretability analyses based on balanced performance observed in cross-validated out-of-fold outputs from the training set. Several non-linear models yielded comparably high discrimination, with overlapping 95% confidence intervals for AUROC.The cross-validated AUC for the XGBoost model in the training set was 0.9932, with a classification threshold of 0.57 derived from the Youden index. In the test set, the AUROC was 0.998 (95% CI: 0.993–1.000), AUPRC was 0.996 (95% CI: 0.988–1.000), accuracy was 0.991 (95% CI: 0.978–1.000), sensitivity was 0.969 (95% CI: 0.930–1.000), specificity was 1.000 (95% CI: 1.000–1.000), precision was 1.000 (95% CI: 1.000–1.000), and F1-score was 0.984 (95% CI: 0.964–1.000). The Brier score for the test set was 0.013. Calibration curves and decision-curve analysis demonstrated favourable probability calibration and higher estimated net benefit across a range of threshold probabilities. SHAP-value analysis highlighted that postoperative-day-1 CRP, postoperative-day-3 FIB, postoperative-day-3 APTT, postoperative-day-3 CRP, postoperative-day-3 FDP, postoperative-day-3 PT, postoperative-day-1 UREA, and postoperative-day-1 CREA were major contributing features. Conclusion The XGBoost model demonstrated excellent discrimination and favorable calibration in the internal test set for identifying postoperative-day-3 ultrasound-detected DVT among patients with FAK. However, several predictors in the final model were measured on postoperative day 3, temporally proximate to the time-point of DVT ascertainment; these biomarkers may reflect biochemical changes secondary to already-formed thrombus rather than pre-thrombotic prospective risk signals. Therefore, the near-perfect discrimination observed in this retrospective cohort should be interpreted cautiously and should not be regarded as definitive evidence of prospective early-prediction performance. Temporal and multicenter external validation using a prespecified prediction time point, ideally restricted to variables available prior to DVT assessment, is required to determine the model's generalizability and prospective clinical applicability.