Prediction models for intraoperative blood loss in spine surgery: a systematic review and methodological appraisal

Abstract

Prediction models may support perioperative blood management in spine surgery, but their credibility depends on outcome definition, prediction time, calibration, validation, and transparent reporting rather than discrimination alone. We searched nine databases from inception to April 1, 2026. Eligible reports developed or independently validated a multivariable model for directly recorded or study-defined estimated quantitative intraoperative blood loss (IBL) or a thresholded IBL event. Transfusion-only, calculated hidden-loss, and combined perioperative/postoperative-loss outcomes were excluded. Poor model reporting alone did not determine eligibility. Two authors independently verified eligibility, report-to-study linkage, extraction, and PROBAST assessment. All eligible model configurations were extracted using CHARMS. Configurations sharing the same participants, outcome, preprocessing, feature-selection, and validation pipeline were treated as one PROBAST assessment family; configuration-specific differences were recorded separately. Validation was classified as random split, temporal, geographic, or fully independent external validation. Sensitivity analyses addressed publication status, outcome measurement, and prediction time. Fourteen reports representing 12 retrospective studies and 84 model configurations were included. Seven studies used conventional statistical methods and five used machine-learning methods. Six studies reported calibration and six reported decision-curve analysis; these were distinguished from other post hoc clinical analyses. Three development reports contained geographic validation cohorts, but only the Shi model had a fully independent multinational companion validation, where discrimination and calibration deteriorated substantially. All studies had high overall PROBAST risk of bias, driven mainly by study-specific analysis limitations rather than retrospective design alone. No prospective model-guided impact study was identified. Current IBL models are not ready for unqualified routine implementation. Evidence is retrospective and heterogeneous, analysis-domain bias is common, independent validation is sparse, and prospective clinical impact is absent. No inference about superiority of statistical or machine-learning approaches is supported.