Artificial Intelligence-Based Prediction of Lower Extremity Deep Vein Thrombosis Risk After Knee/Hip Arthroplasty

Abstract

Deep vein thrombosis (DVT) is a common postoperative complication of knee/hip arthroplasty. There is a continued need for artificial intelligence-based methods of predicting lower extremity DVT risk after knee/hip arthroplasty. In this study, we performed a retrospective study to analyse the data from patients who underwent primary knee/hip arthroplasty between January 2017 and December 2021 with postoperative bilateral lower extremity venous ultrasonography. Patients' features were extracted from electronic health records (EHRs) and assigned to the training (80%) and test (20%) datasets using six models: eXtreme gradient boosting, random forest, support vector machines, logistic regression, ensemble, and backpropagation neural network. The Caprini score was calculated according to the Caprini score measurement scale, and the corresponding optimal cut-off Caprini score was calculated according to the largest Youden index. In total, 6897 cases of knee/hip arthroplasty were included (average age, 65.5 ± 8.9 years; 1702 men), among which 1161 (16.8%) were positive and 5736 (83.2%) were negative for deep vein thrombosis. Among the six models, the ensemble model had the highest area under the curve [0.9206 (0.8956, 0.9364)], with a sensitivity, specificity, positive predictive value, negative predictive value, and F1 score of 0.8027, 0.9059, 0.6100, 0.9573 and 0.7003, respectively. The corresponding optimal cut-off Caprini score was 10, with an area under the curve, sensitivity, specificity, positive predictive value, and negative predictive values of 0.5703, 0.8915, 0.2491, 0.1937, 0.9191, and 0.3183, respectively. In conclusion, machine learning models based on EHRs can help predict the risk of deep vein thrombosis after knee/hip arthroplasty.

Keywords: arthroplasty; artificial intelligence; deep vein thrombosis.

Figure 1.

Free-text extraction and labelling.

Figure 2.

Receiver operating characteristic curves of the six models and Caprini score. Abbreviations: ROC, receiver operating characteristic curve; AUC, area under the receiver operating characteristic curve; SVM, support vector machines; LR, logistic regression; RF, random forest; BPNN, backpropagation neural network; XGBoost, eXtreme gradient boosting.