Eur Spine J. 2026 Nov 15. doi: 10.1007/s00586-026-10459-5. Online ahead of print.
ABSTRACT
BACKGROUND: Postoperative outcomes following cervical spine surgery remain heterogeneous, limiting effective preoperative risk stratification. Conventional prognostic approaches rely on demographic, clinical, and qualitative imaging factors but incompletely capture complex, high-dimensional information present in modern imaging. Machine learning-based imaging models have emerged as promising tools for individualized outcome prediction; however, the scope, performance, and methodological quality of this literature have not been systematically synthesized.
METHODS: A systematic review was conducted in accordance with PRISMA 2020 guidelines. PubMed and Scopus were searched from inception through November 2025 for English-language studies applying machine learning or artificial intelligence techniques to imaging data to predict postoperative outcomes following cervical spine surgery. Data on study design, imaging modality, model architecture, validation strategy, performance metrics, and risk of bias (PROBAST) were extracted and synthesized narratively.
RESULTS: The search identified 117 records, of which 16 studies published between 2021 and 2025 met inclusion criteria after full-text or detailed eligibility review. Studies addressed prediction of functional or neurologic recovery (7 studies), postoperative complications (4 studies), adjacent segment degeneration or disease (2 studies), utilization or biomechanical recovery metrics (2 studies), and fusion status assessment (1 study). Sample sizes ranged from 47 to 945 patients. Most studies used internal validation only, reported discrimination metrics, and demonstrated high or unclear risk of bias on PROBAST, primarily within the analysis domain.
CONCLUSIONS: Machine learning-based imaging models show promise for improving prediction of postoperative outcomes following cervical spine surgery, particularly for neurologic recovery and complication risk. However, most existing models rely on internal validation and demonstrate methodological limitations that restrict clinical translation. Rigorous external validation, standardized imaging workflows, and improved reporting of calibration and clinical utility are needed before these approaches can be responsibly integrated into routine practice.
PMID:43034949 | DOI:10.1007/s00586-026-10459-5