Background: Venous thromboembolism (VTE) following total hip arthroplasty and total knee arthroplasty (TKA) is linked to immobility, and preoperative prediction remains difficult. We aimed to evaluate whether annual mean length of stay (LOS) is associated with the incidence of VTE and develop a generalizable machine learning model to preoperatively predict the incidence of symptomatic VTE following total hip and TKA using National Surgical Quality Improvement Program.
Methods: Annual incidence of 30-day postoperative VTE, deep vein thrombosis, and pulmonary embolism was calculated over 6 years and tested for trend. Correlation between annual VTE rates and mean LOS was calculated. Predictive models (logistic regression, random forest, and XGBoost) were trained and tested based on year of surgery with different oversampling algorithms used to address data imbalance.
Results: A total of 498,314 patients were included, with 0.88% developing a VTE within 30 days. VTE rates decreased from 1.11% in 2014 to 0.76% in 2019 (P < .001). There was a strong correlation between the yearly incidence of VTE, pulmonary embolism, and deep vein thrombosis and mean LOS (r = 0.96, 0.87, and 0.98, respectively). Univariate analysis demonstrated that TKA, inpatient setting, American Society of Anesthesiologists classification, and various patient comorbidities were significantly associated with VTE. The logistic regression model trained on all data with a balanced loss scoring function performed the best (area under the curve = 0.600).
Conclusions: This study revealed declining VTE rates strongly correlated to decreasing postoperative LOS and identified patient and surgery-specific factors associated with VTE risk. Development of more accurate machine learning models for VTE prediction may improve risk stratification, prevention, and monitoring for arthroplasty patients.
Keywords: Machine learning; Outcome prediction; Total hip arthroplasty; Total knee arthroplasty; Venous thromboembolism.
© 2024 The Authors.