Development of a preoperative prediction tool for massive intraoperative blood loss in spinal metastases surgery...

Clinical models for predicting massive intraoperative blood loss (IBL) in spinal metastasis surgery exhibit a systematic, vascularity-dependent bias, underestimating risk in non-hypervascular tumors while overestimating it in hypervascular ones. We aimed to develop and validate an AI model integrating magnetic resonance imaging (MRI) radiomics to reduce this bias and improve risk stratification.

This retrospective study included 601 patients who underwent surgery for spinal metastases between January 2016 and December 2022. They were randomized to a development cohort ( n = 479) and a test cohort ( n = 122). Clinical characteristics and radiomic features from T1c MRI were used to develop predictive models. Based on internal validation across nine machine learning algorithms, the best-performing model was selected. External testing was performed using an independent cohort of 101 patients to assess generalizability. The primary outcome was defined as massive IBL, with an estimated blood loss of 2500 mL or more. Model performance was evaluated using the area under the curve (AUC), calibration curves, and decision curve analysis. An artificial intelligence (AI) tool was developed to facilitate clinical use.

Among the 702 patients included, the combined model integrating MRI radiomics and clinical variables outperformed the clinical model in both internal [AUC: 0.901 (95% CI: 0.8330-0.9690) vs. 0.735 (95% CI: 0.6238-0.8458)] and external validation cohorts [AUC: 0.885 (95% CI: 0.8052-0.9639) vs. 0.604 (95% CI: 0.4355-0.7720)]. Subgroup analysis revealed that in non-hypervascular tumors, the combined model significantly increased the sensitivity for identifying massive bleeding (0.85 vs. 0.30, P < 0.001). In hypervascular tumors, the specificity was notably enhanced (0.81 vs. 0.55, P < 0.001), and meanwhile the false-positive rate was reduced. The use of AI tools also improved the prediction performance of spine surgeons. The model is freely accessible for download at https://github.com/banluqihao/A-predict-tool-for-spinal-metastases-surgery .

By integrating MRI radiomics features, our model reduces the systemic biases of clinical-only models that depend on unreliable histological surrogates. This enables more accurate and individualized risk stratification, providing a reliable tool to guide preoperative planning and support more accurate risk stratification for patients with spinal metastases.