Prediction models for progression and outcomes in degenerative cervical myelopathy: a systematic review and critical appraisal.

OBJECTIVE: This systematic review evaluated the methodological quality, predictive performance, risk of bias, and clinical applicability of prediction models for disease progression and postoperative outcomes in degenerative cervical myelopathy (DCM), aiming to provide evidence for model optimization and clinical translation. METHODS: PubMed, Embase, and Web of Science were searched from inception to May 2026. Original studies developing or validating DCM prediction models were included. Data extraction followed CHARMS, risk of bias was assessed using PROBAST, and reporting quality was evaluated with TRIPOD. RESULTS: Twelve studies (prospective and retrospective cohorts from China, the US, Canada, and the Czech Republic) were included. Modeling approaches ranged from conventional regression to machine learning algorithms, with AUCs from 0.53 to 0.93. Major predictors included age, baseline JOA/mJOA, symptom duration, smoking, comorbidities, and radiomic features. PROBAST revealed low bias in participants, predictors, and outcomes for most studies, except Kadanka et al. (2017) which showed high outcome bias; however, all studies had high analysis-domain bias due to inadequate sample-event matching, inappropriate continuous variable handling, inconsistent missing data management, and insufficient calibration reporting. Only five studies reported both discrimination and calibration, and four completed both internal and external validation. TRIPOD compliance ranged from 59.1 to 81.8%, with sample size planning, calibration reporting, and model coefficients being most underreported. Clinical translation remains limited. CONCLUSION: Current DCM prediction models carry high overall bias and methodological heterogeneity. Parsimonious clinical models using readily available predictors (age, baseline mJOA, symptom duration) with rigorous validation offer advantages in calibration and interpretability over high-dimensional radiomic machine learning models. Future research should prioritize external validation of simple models and development of bedside-friendly tools, with adherence to TRIPOD and PROBAST standards. SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/PROSPERO/view/CRD420261407781, CRD420261407781.