Development, internal testing, and external validation of an interpretable machine-learning model for predicting imaging-confirmed in-hospital postoperative deep vein thrombosis in older patients with patellar fractures.

backgroundOlder patients with patellar fractures may be at increased risk of postoperative deep vein thrombosis (DVT) because of trauma, perioperative immobilization, and age-related prothrombotic susceptibility. However, risk-stratification tools tailored to this specific population remain limited. We aimed to develop, internally test, and externally validate an interpretable machine-learning model, based on routinely available early admission data, for predicting imaging-confirmed in-hospital postoperative DVT in older patients with patellar fractures.

methodsThis retrospective prediction-model study included an internal cohort of 741 patients aged ≥ 65 years who underwent surgery for patellar fractures between January 2017 and December 2022. The internal cohort was randomly divided into a development set (n = 518) and a held-out internal test set (n = 223). An independent external validation cohort of 273 eligible patients from a separate tertiary hospital was used to evaluate model transportability. A three-step feature-selection pipeline combining variance inflation factor screening, LASSO regression, and recursive feature elimination with cross-validation was used to identify core predictors from routine admission variables. Seven machine-learning algorithms, including reference linear benchmark models, were developed and compared. The final model was externally validated and compared with the preoperative Caprini score as an established clinical VTE risk-assessment benchmark. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC), Brier score, calibration analysis, decision-curve analysis, and bootstrap optimism correction. SHapley Additive exPlanations were used to describe the behavior of the final model.

resultsIn the internal cohort, the overall incidence of imaging-confirmed in-hospital postoperative DVT was 18.5% (137/741). In the external validation cohort, 38 of 273 patients developed DVT. Nine predictors were retained: albumin, platelet count, fibrinogen, sodium, body mass index, total protein, cholinesterase, low-density lipoprotein cholesterol, and hemoglobin. Among the candidate models, XGBoost achieved the highest held-out internal AUROC of 0.927 (95% CI, 0.873-0.969), which was interpreted as internal testing rather than independent validation. Bootstrap optimism correction yielded an optimism-corrected AUROC of 0.904 and an optimism-corrected Brier score of 0.089. In the external validation cohort, the XGBoost model achieved an AUROC of 0.838 (95% CI, 0.758-0.904) and a Brier score of 0.081. Compared with the preoperative Caprini score, XGBoost showed higher AUROC in both the internal test set (0.927 vs. 0.687) and the external validation cohort (0.838 vs. 0.685), with lower Brier scores in both settings. Descriptive analysis of thrombus location showed that distal DVT was the predominant classifiable subtype in both the internal and external cohorts. SHAP analysis indicated that albumin, platelet count, fibrinogen, body mass index, and sodium were among the most influential predictors, but these patterns were interpreted as model-derived associations rather than causal effects or clinical thresholds.

conclusionWe developed, internally tested, and externally validated an interpretable XGBoost model based on nine routine early admission variables for predicting imaging-confirmed in-hospital postoperative DVT in older patients with patellar fractures. The model showed higher discrimination and lower prediction error than the preoperative Caprini score in both validation settings; however, because the endpoint was an imaging-ascertained composite outcome and detailed surgical and perioperative management variables were not fully captured, residual perioperative-management confounding cannot be excluded. Therefore, the model should be considered exploratory. Clinically, a high predicted risk should be interpreted as a prompt for closer postoperative surveillance and individualized clinical review, including consideration of repeat duplex ultrasonography when appropriate, verification of thromboprophylaxis adherence, hydration optimization, and reassessment of modifiable risk factors. The model should not be used alone to intensify or extend anticoagulation, change rehabilitation protocols, alter fluid management, or determine other postoperative management decisions. Further multicenter validation, recalibration, and evaluation of clinically meaningful DVT subtypes are required before clinical implementation.