Significant blood loss (≥ 500 mL) during spinal surgery is common and linked to greater transfusion requirements and worse outcomes. Preoperative tools for risk stratification and individualized blood-management strategies are lacking. We aimed to develop a machine-learning model predicting intraoperative significant blood loss and to assess whether a risk-adapted blood management improves clinical benefit.

We used data from 3944 spinal surgery patients at Peking Union Medical College Hospital (December 2018–October 2021) to train 26 machine-learning algorithms. We used Shapley additive explanations to identify key predictors from 49 candidate variables to train simplified models. The optimal simplified model was externally validated in 843 patients from West China Hospital. Decision-curve analysis and spline analysis were used to evaluate Cell Saver benefit across model-predicted risk.

The 12-variable ranger model achieved an AUC of 0.814 (95% CI, 0.790–0.839) in the test set and 0.820 (0.785–0.854) in the external cohort. Decision-curve analysis demonstrated that the risk-adapted Cell Saver strategy provided greater net benefit than current practice. Spline analysis demonstrated that Cell Saver benefit rose with increasing predicted risk: for risk >0.53, Cell Saver use was associated with higher postoperative hemoglobin; for risk >0.58, it reduced allogeneic transfusion requirements. Among patients requiring allogeneic transfusion, Cell Saver use decreased red-cell unit volume at all risk levels, with larger reductions in higher-risk patients.

This 12-variable machine-learning model can accurately predict significant blood loss risk in spinal surgery. Risk-adapted Cell Saver use guided by predicted risk provides greater net clinical benefit than experienced-based real-world strategy.