Development of Machine Learning Algorithms for Predicting Preoperative and Postoperative venous Thromboembolism in Patients Undergoing Surgery for Spinal Metastasis

Abstract

Objective: This study aims to develop and compare machine learning models (MLMs) for predicting venous thromboembolism (VTE) in patients undergoing surgery for spinal metastasis. The study evaluates the predictive capabilities of MLMs for preoperative and postoperative VTE within different time frames. Materials and Methods: A total of 334 patients undergoing surgery for spinal metastasis were included, with a mean age of 57.6 years and 57.2% being male. The investigation assessed postoperative VTE prevalence within 30 and 90 days, with pulmonary embolism (PE) and deep vein thrombosis (DVT) rates at 20% and 80%, respectively. Key patient-related factors—age, body mass index, preoperative ambulatory status, albumin level, hemoglobin level, partial thromboplastin time, and operative time—were considered potential predictors of VTE. Results: The postoperative VTE prevalence was 8.98% within 30 days and 13.47% within 90 days. Age, body mass index, preoperative ambulatory status, albumin level, hemoglobin level, partial thromboplastin time, and operative time emerged as significant VTE predictors. The gradient boosted tree algorithm was the best-performing MLM for predicting VTE within 90 days, with AUC values of 0.77 preoperatively and 0.71 postoperatively. For predicting VTE within 30 days, the support vector machine model was most effective, with AUCs of 0.72 preoperatively and 0.68 postoperatively. Conclusion: Predictive analytics and MLMs effectively predict preoperative and postoperative VTE in patients undergoing surgery for spinal metastasis. Identified key factors and MLM performance metrics offer valuable insights for risk assessment and preventive measures in this patient population.