To systematically evaluate the application of machine learning (ML)-based predictive models in assessing the risk of refractures following vertebral augmentation (VA) in patients with osteoporotic vertebral compression fractures (OVCFs). The methodology and performance of these models will be assessed, and key predictors will be identified. This will provide a reference for constructing high-quality predictive models in the future. A systematic search was conducted to identify relevant studies using eight electronic databases for the time period from database inception to February 28, 2026. Two reviewers independently conducted literature searches, screening, and data extraction. Eligible studies include all retrospective or prospective studies of ML models that have been developed or validated in patients with OVCFs and that predict refracture after VA. Model bias and applicability were evaluated using the Predictive Model Risk of Bias Assessment Tool (PROBAST). The quality of transparent reporting was assessed using the Multivariate Predictive Models for Individual Prognosis or Diagnosis-Artificial Intelligence (TRIPOD + AI). A total of 14 studies were included, involving 73 postoperative refractures prediction models for patients with OVCFs, constructed using 25 ML algorithms. The incidence of refractures ranged from 10.6% to 64.4%. The high-frequency predictors network diagram showed that 13 appeared at least twice and more, with age, body mass index, and radiomics features being the most common predictors. Methodological appraisal using PROBAST revealed that all 14 studies exhibited a high risk of bias (ROB). The primary risks of bias were identified as follows: a retrospective study design, a small sample size, the failure to report missing data handling methods, inappropriate variable selection methods, and a lack of external validation. The compliance of the included studies with each item on the TRIPOD + AI checklist also needs to be improved. The aspects where the quality of reports is generally insufficient include: missing sample size calculation, opaque processing of missing data, failure to provide a complete model, absence of research-related protocols and registration information, and lack of clinical application considerations. Our results indicated that most published predictive modelling studies exhibited robust predictive performance, but they still demonstrated methodological shortcomings and a high ROB. Research on predictive models for postoperative refractures risk in patients with OVCFs undergoing VA is still in its developmental stage. Future studies should adhere strictly to predictive model reporting standards and enhance validation processes to enable more effective clinical application. This study is registered with PROSPERO (Registration ID: CRD 420251082219).
Shi et al. (Sat,) studied this question.