Use of Machine Learning and Imaging to Predict Outcomes After Lumbar Spine Surgery: A Systematic Review

Postoperative outcomes following lumbar spine surgery remain variable, and existing clinical risk stratification methods offer limited predictive accuracy. Advances in machine learning enable quantitative analysis of routinely acquired spinal imaging, creating opportunities to improve preoperative outcome prediction. However, the methodological quality and clinical readiness of imaging-based machine learning models in lumbar spine surgery have not been systematically evaluated. This systematic review followed PRISMA 2020 guidelines. PubMed, Embase, and Scopus were searched through October 2025 for studies using machine learning models trained on imaging data (MRI, CT, or radiographs) to predict postoperative outcomes after lumbar spine surgery. Two reviewers independently screened studies, extracted data, and assessed risk of bias using the Prediction model Risk Of Bias ASsessment Tool (PROBAST). Twenty-six studies published between 2019 and 2025 met inclusion criteria. Most models predicted patient-reported or clinical outcomes, with fewer addressing complications, structural outcomes, or imaging-derived biomarkers. MRI was the predominant modality. Model discrimination was frequently moderate to high (median AUC, 0.83), though external validation and calibration reporting were uncommon. All studies demonstrated high risk of bias, primarily related to analysis and validation practices. Imaging-based machine learning models show promise for improving preoperative outcome prediction after lumbar spine surgery, but methodological limitations limit current clinical applicability. Greater emphasis on external validation, standardized outcomes, and prospective evaluation is needed to support safe translation into routine spine care.