Purpose To synthesise and rank machine learning (ML) and traditional methods for postoperative venous thromboembolism (VTE) prediction using Bayesian network meta-analysis, with Caprini as the reference. Methodology Medline (OVID), Embase, and the Cochrane Library were searched for studies directly comparing ML with traditional risk models for predicting postoperative VTE. Risk of bias was assessed with PROBAST+AI. The primary outcome was model discrimination measured by the area under the receiver operating characteristic (AUROC) curve. A Bayesian random effects network meta-analysis estimated mean differences (MD) in AUROC vs Caprini with 95% credible intervals (CrIs). Best and worst version analyses addressed within study multiplicity from multiple candidate models within the same ML class. Subgroup analyses examined cancer surgery and procedure type. Results Eighteen studies were included. In the best version analysis, metaensembles demonstrated the largest improvement over Caprini (MD = 0.231, 95% CrI (0.117, 0.346)); multiple ML classes also improved, whereas most traditional scores and surgeon judgement did not differ significantly. Findings were consistent in the worst version analysis. SUCRA rankings placed metaensembles first (SUCRA 88.89) and Caprini 14th (SUCRA 28.57), with Geneva lowest (SUCRA 3.32). In cancer surgery cohorts, most models did not differ significantly from Caprini; metaensembles was borderline in the best version analysis and statistically significant in the worst version (MD = 0.222, 95% CrI (0.0121, 0.432)). Surgery type subgroup analyses showed wider uncertainty than the overall analysis. Conclusion Metaensemble models demonstrated the strongest discrimination and outperformed traditional risk scores overall. Although EHR embedded decision support could enhance stratification, translation to routine care remains uncertain due to methodological and validation gaps. Further prospective, multicentre implementation studies are needed.