Construction of a risk prediction model for postoperative deep vein thrombosis(DVT)in pa-tients with traumatic lower limb fractures based on gradient boosting decision tree

Liu Ruiting; Xie Suli; Wang Ke; Jiao Wencang; Zhao Yan; Chen Lili; Zheng Xican

doi:10.3969/j.issn.1009-4237.2025.07.008

Abstract：Objective To develop a deep vein thrombosis(DVT)risk prediction model using the gradient boosting decision tree(GBDT)algorithm among patients with traumatic lower limb fractures,and to validate the mod-el by comparing its performance with the risk assessment profile for thromboembolism(RAPT)scale.Methods A retrospective study was performed to recruit 591 patients with traumatic lower limb fractures admitted to No.988 Hospital of the PLA Joint Logistics Support Force between Jan.2020 and Dec.2023.This cohort included 333 males and 258 females,aged 18-99(mean 58.9)years.The dataset was randomly split into a training set(445 ca-ses)and an internal validation set(146 cases)at a 3∶1 ratio.Another 76 patients with traumatic lower limb frac-tures from two additional hospitals(Zhengzhou Orthopedics Hospital and Huixian People's Hospital)between Jan.and Mar.2024 were prospectively enrolled for external validation.Using the results of univariate analysis and the GBDT algorithm,a DVT risk prediction model was developed based on the training set data,whose predictive per-formance was assessed among the internal and external validation sets for sensitivity,specificity,area under the curve(AUC),and accuracy.Additionally,its performance was compared with that of the RAPT scale.Results DVT oc-curred in 119 patients(20.1%,95%CI:16.9%-23.4%).Significant differences were observed in age,body mass index(BMI),fracture type,injury mechanism,injury energy,anesthesia method,and D-dimer levels between the thrombosis and non-thrombosis groups(all P<0.05).The GBDT algorithm identified fracture type,age,BMI,and in-jury mechanisms as significant predictors of postoperative DVT in patients with traumatic lower limb fractures.Addi-tionally,the model stratified patients into four high-risk subgroups based on the abovementioned predictors:(1)frac-ture type of pelvic fracture+age≥63 years;(2)fracture type not of pelvic fracture+age≥80 years+BMI≥22 kg/m2+cause of injury of road traffic accident or fall from height;(3)fracture type of tibiofibular or femoral fracture+36≤age<80 years+BMI of(24-25)kg/m2+cause of injury of road traffic accident or fall from height;and(4)fracture type of tibiofibular or femoral fractures+64≤age<80 years+BMI≥25 kg/m2+cause of injury of road traffic accident or fall from height.Receiver operating characteristic(ROC)curve analysis showed that the AUC of the GBDT algo-rithm was 0.772(95%CI:0.718-0.826)in the training set,0.785(95%CI:0.703-0.867)in the internal validation set,and 0.743(95%CI:0.546-0.939)in the external validation set.No statistically significant difference was ob-served among the AUC values(P=0.796).The AUC of the RAPT scale in the training set was 0.626(95%CI:0.558-0.694),which was significantly lower than 0.772(95%CI:0.718-0.826)of the GBDT algorithm(P<0.001).Conclusion The GBDT-based DVT risk prediction model developed in this study exhibits good predictive performance and can be used for early identification of postoperative DVT risk in patients with traumatic lower limb fractures.