Accuracy of machine learning in predicting the risk of venous thromboembolism after spine surgery: a systematic review and meta-analysis

Abstract Background Venous thromboembolism (VTE) is a common and serious complication after spine surgery, increasing hospital stays, medical costs, and mortality risk. Machine learning (ML) models are increasingly being used for postoperative VTE prediction; however, their performance varies widely due to differences in variable selection and algorithms. This meta-analysis systematically evaluated the predictive performance of ML models for VTE after spine surgery. Methods This predictive model-focused meta-analysis was prospectively registered with PROSPERO (CRD420261412003) and reported using PRISMA 2020 guidelines. PubMed, Web of Science, Medline, and Embase databases were searched from their inception to June 1, 2026. The risk of bias among the eligible studies was evaluated using the PROBAST tool, and the certainty of the pooled evidence was assessed using the GRADE framework. The concordance index (C‑index) was used as the primary measure of discrimination, and a random-effects model was applied for quantitative data synthesis. Results The initial search yielded 579 articles, of which 25 studies encompassing 105 ML models met the inclusion criteria. The pooled C-index was 0.823 (95% confidence interval [CI] 0.783–0.864) for the training cohorts and 0.778 (95% CI 0.732–0.824) for the validation cohorts. Subgroup analyses revealed relatively superior model stability in the metastatic tumor and fracture surgery subgroups, with the metastatic tumor subgroup attaining moderate GRADE evidence certainty, while the other subgroups exhibited low or very low certainty. In contrast, substantial heterogeneity was observed in the fusion surgery subgroup. Due to the limited number of studies, robust inferences could not be drawn for the infection and deformity surgery subgroups. Conclusions Current evidence supports the exploratory potential of ML for predicting VTE after spine surgery and aiding postoperative risk stratification. However, the clinical adoption of the findings remains limited by high methodological heterogeneity, insufficient external validation, and risk of bias. Future studies should prioritize multicenter prospective data, standardized modeling, and rigorous external validation to confirm clinical utility.