Risk prediction in spine surgery: a scoping review of traditional models, artificial intelligence, and the challenge of clinical translation

BACKGROUND: Accurate perioperative risk stratification is essential to patient safety and informed consent in spine surgery. Traditional regression-based risk scores are widely used but have modest predictive performance, while artificial intelligence and machine learning (AI/ML) approaches are increasingly applied to surgical risk prediction. This review synthesizes evidence on spine surgery risk prediction models, compares traditional approaches with AI/ML methods, and identifies key translational barriers. METHODS: A structured literature search was conducted across major biomedical databases using combinations of spine surgery, risk prediction, perioperative outcomes, and AI-related terms, with additional conceptual searches targeting explainability, validation, and clinical translation. Studies were selected for relevance to adult spine surgery risk prediction, model development or validation, and methodological or translational considerations. Given substantial heterogeneity in study design and outcomes, findings were synthesized qualitatively using a narrative approach. RESULTS: Traditional spine-specific risk models demonstrate fair-to-good discrimination for common outcomes, with typical areas under the receiver operating characteristic curve (AUCs) ranging from approximately 0.64 to 0.78. AI/ML models often report modest improvements in discrimination over regression-based approaches, particularly for common outcomes such as intensive care unit (ICU) admission and mortality, but gains are inconsistent and context-dependent. Across both model types, external validation, calibration drift, limited prospective outcome evidence, and challenges related to interpretability and workflow integration remain prominent. CONCLUSIONS: Traditional risk models remain interpretable, trusted, and competitively performant for many spine surgery outcomes. While AI/ML approaches expand data integration and interaction modeling, their clinical impact is constrained by validation, trust, and implementation barriers. Future progress will depend less on incremental performance gains and more on rigorous external validation, prospective outcome studies, and integration into clinical workflows.